CN120496055B - A Vision-Based Method and System for Quality Inspection of Paper Tray Production - Google Patents

A Vision-Based Method and System for Quality Inspection of Paper Tray Production

Info

Publication number
CN120496055B
CN120496055B CN202510710915.XA CN202510710915A CN120496055B CN 120496055 B CN120496055 B CN 120496055B CN 202510710915 A CN202510710915 A CN 202510710915A CN 120496055 B CN120496055 B CN 120496055B
Authority
CN
China
Prior art keywords
map
image
residual
paper tray
mold
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202510710915.XA
Other languages
Chinese (zh)
Other versions
CN120496055A (en
Inventor
叶锦强
胡可一
黄陈其
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Qingyuan Huayan New Material Technology Co ltd
Qingyuan Yuchen Environmental Protection New Materials Co ltd
Qingyuan Keding Electromechanical Equipment Co ltd
Original Assignee
Qingyuan Huayan New Material Technology Co ltd
Qingyuan Yuchen Environmental Protection New Materials Co ltd
Qingyuan Keding Electromechanical Equipment Co ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Qingyuan Huayan New Material Technology Co ltd, Qingyuan Yuchen Environmental Protection New Materials Co ltd, Qingyuan Keding Electromechanical Equipment Co ltd filed Critical Qingyuan Huayan New Material Technology Co ltd
Priority to CN202510710915.XA priority Critical patent/CN120496055B/en
Publication of CN120496055A publication Critical patent/CN120496055A/en
Application granted granted Critical
Publication of CN120496055B publication Critical patent/CN120496055B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • G06V10/454—Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/64—Three-dimensional [3D] objects
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/70—Labelling scene content, e.g. deriving syntactic or semantic representations
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
    • G01N21/84—Systems specially adapted for particular applications
    • G01N21/88—Investigating the presence of flaws or contamination
    • G01N21/8851—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges
    • G01N2021/8887—Scan or image signal processing specially adapted therefor, e.g. for scan signal adjustment, for detecting different kinds of defects, for compensating for structures, markings, edges based on image processing techniques
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Signal Processing (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a paper holder production quality detection method and a system based on visual processing, which belong to the technical field of quality detection and comprise the steps of collecting images and motion data, aligning according to a time sequence, classifying risk categories, marking high risk areas, constructing a stress prediction distribution diagram, calculating angle residual errors according to the existing images and the stress prediction distribution diagram, calculating an angle residual error diagram to obtain an area average residual error value, taking the area average residual error value as a weighting coefficient, generating a speed and angular speed regulating factor by a deviation distribution index, and finally generating a next period mould execution instruction based on an adjustment strategy. According to the detection method, the process modeling and the structure sensing are combined, so that the defects of the traditional paper support defect detection means based on the static image in the aspects of time lag, structure invisibility, control unreachablility and the like are effectively overcome, and the intelligent level and the detection coverage rate of paper support quality detection are remarkably improved.

Description

Paper support production quality detection method and system based on visual processing
Technical Field
The invention belongs to the technical field of quality detection, and particularly relates to a paper support production quality detection method and system based on visual processing.
Background
Along with the popularization of the green packaging and environmental protection concepts, the paper holder is widely applied to product packaging in a plurality of industries such as consumer electronics, medicines, foods and the like due to good buffering property, biodegradability and cost advantages. The forming process of the paper tray generally comprises the key steps of pulping, injection molding, compression molding, demolding and the like, wherein the demolding link has the greatest influence on the quality of the final product. In actual production, the paper support often causes quality problems such as collapse, cracks, fiber layer breakage, edge wrinkling and the like due to improper operation in the demolding process, unreasonable setting of mold parameters or fluctuation of pulp state. These defects not only affect the physical strength and structural stability of the product, but also pose an unpredictable risk of damage during subsequent packaging, shipping and use.
The paper support quality detection method commonly adopted in the current industry is mainly based on the static image acquisition and processing technology after the forming is finished, and defect identification is carried out by means of a visual camera system and low-level features such as edges and textures. However, there are several significant disadvantages to this type of approach. First, some structural defects of the paper tray, such as insufficient local fiber compaction, initial collapse tendency, etc., appear inconspicuous on the surface after the formation is completed, and are difficult to identify by conventional vision systems. Secondly, most of the prior art relies on surface images for discrimination, and the defect 'cause' is lack of understanding, especially dynamic behavior abnormality formed in the paper tray demolding process cannot be captured and modeled, so that the detection result has hysteresis and high false detection rate. In addition, the traditional system is only responsible for 'detection' function, lacks feedback capability to equipment operation parameters, and cannot participate in production control closed loop, so that real-time prediction and adjustment to quality problems cannot be realized.
Therefore, we propose a paper support production quality detection method and system based on visual processing to solve the above problems.
Disclosure of Invention
The invention aims to solve the problem that the real-time prediction and adjustment of quality problems cannot be realized in the prior art, and provides a paper support production quality detection method and system based on visual processing.
In order to achieve the above purpose, the present invention adopts the following technical scheme:
a paper support production quality detection method based on visual processing comprises the following steps:
The method comprises the steps of collecting original data, aligning an image sequence with motion parameters according to time points and outputting a behavior data sequence, wherein the original data comprises the image sequence and the corresponding mold motion parameters;
Extracting space-time convolution characteristics, motion trend characteristics and state change trends from a behavior data sequence, splicing to obtain a final characterization vector, classifying risk categories of the final characterization vector, outputting risk category labels of the final characterization vector, and simultaneously carrying out weighted summation on a final layer of convolution activation diagram of the final characterization vector to obtain a space abnormality significance diagram;
Splicing a final image in a demolding process with a spatial anomaly significance map, and constructing a stress prediction distribution map by combining risk category labels, wherein the training in the prediction process adopts a multi-objective loss function, the multi-objective training function comprises a mean square error loss term weighted by a significant region, a structure edge preservation regular term and a structure continuity consistency regular term, and the structure continuity consistency regular term is obtained by enabling the gradient directions of the final image and the stress prediction distribution map to be consistent;
Collecting a fiber structure diagram, extracting a local main structure directional diagram from the fiber structure diagram, extracting a gradient directional diagram from a stress prediction distribution diagram, calculating an angle residual diagram between the local main structure direction and the gradient directional diagram, and generating a deviation distribution index by combining the angle residual diagram and regional weights, wherein the regional weights are obtained through a space anomaly significance diagram;
And calculating an average residual value of the region from the angle residual map, taking the average residual value as a weighting coefficient and a deviation distribution index generation speed and angular speed adjusting factor, and finally generating a next period mould executing instruction based on an adjusting strategy.
Preferably, the die movement parameters include a linear velocity of the die movement and a release angle between the die and the paper holder.
Preferably, the motion trend feature is obtained by processing a corresponding optical flow sequence, wherein the optical flow sequence is obtained by combining inter-frame motion vector fields, and the inter-frame motion vector fields are obtained by inputting adjacent frame images in the behavior data sequence into an optical flow estimation network for calculation.
Preferably, the risk category labels include normal, mucosal, stretch-break, and collapse.
Preferably, an abnormal residual regularization term is introduced in the risk category classification training process, the abnormal residual regularization term is generated by extracting average codes from a normal demolding sequence through a reference behavior template which is constructed by counting demolding samples in advance, and L2 residual errors of the current behavior characteristics and the reference template are calculated and added into classification loss to distinguish pseudo-abnormal demolding risks from real demolding risks.
Preferably, the local principal structure pattern is obtained by analyzing the same structure Zhang Liang, calculating the directionality of each pixel within a Gaussian weighted window.
Preferably, the deviation distribution index is obtained by combining a weighted aggregation of the dynamic saliency map and the directional residual map, and normalizing the severity of the predicted deviation of the quantization model.
Preferably, the adjustment strategy comprises:
for a high deviation area of the structure, the speed of the die is reduced, and the pulling impact on the paper holder is slowed down;
and when the integral prediction reliability is not high, the angle transition amplitude of the die is slowed down, and the cracking caused by too fast structural deformation is avoided.
Paper support production quality detecting system based on vision processing includes:
the data acquisition module acquires original data, wherein the original data comprises an image sequence and corresponding mould motion parameters, aligns the image sequence with the motion parameters at a time point and outputs a behavior data sequence;
the anomaly detection module extracts space-time convolution characteristics, motion trend characteristics and state change trends from the behavior data sequence, splices the space-time convolution characteristics, the motion trend characteristics and the state change trends to obtain a final characterization vector, classifies risk categories of the final characterization vector, outputs risk category labels of the final characterization vector, and simultaneously performs weighted summation on a final layer of convolution activation diagram of the final characterization vector to obtain a space anomaly significance diagram;
The stress prediction module is used for splicing a final image of the demolding process with the spatial abnormality significance map and constructing a stress prediction distribution map by combining a risk category label, wherein the training of the prediction process adopts a multi-objective loss function, the multi-objective training function comprises a mean square error loss term weighted by a significant region, a structure edge preservation regular term and a structure continuity consistency regular term, and the structure continuity consistency regular term is obtained by enabling the gradient directions of the final image and the stress prediction distribution map to be consistent;
the image verification module acquires a fiber structure diagram, extracts a local main structure direction diagram from the fiber structure diagram, extracts a gradient direction diagram from a stress prediction distribution diagram, calculates an angle residual diagram between the local main structure direction and the gradient direction diagram, and combines the angle residual diagram and regional weight to generate a deviation distribution index;
And the strategy generation module is used for calculating an area average residual value from the angle residual map, taking the area average residual value as a weighting coefficient and a deviation distribution index generation speed and angular speed regulating factor, and finally generating a next period mould execution instruction based on an adjustment strategy.
In summary, the invention has the technical effects and advantages that the defects of the traditional paper holder defect detection means based on the static image in the aspects of time lag, structural invisibility, control unreachablility and the like are effectively overcome, and the intelligent level and the detection coverage rate of paper holder quality detection are remarkably improved.
Drawings
FIG. 1 is a flow chart of the method steps of the present invention;
fig. 2 is a schematic diagram of a system structure in the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments.
As shown in fig. 1, the paper support production quality detection method based on visual processing includes:
The method comprises the steps of collecting original data, aligning an image sequence with motion parameters according to time points and outputting a behavior data sequence, wherein the original data comprises the image sequence and the corresponding mold motion parameters;
Extracting space-time convolution characteristics, motion trend characteristics and state change trends from a behavior data sequence, splicing to obtain a final characterization vector, classifying risk categories of the final characterization vector, outputting risk category labels of the final characterization vector, and simultaneously carrying out weighted summation on a final layer of convolution activation diagram of the final characterization vector to obtain a space abnormality significance diagram;
Splicing a final image in a demolding process with a spatial anomaly significance map, and constructing a stress prediction distribution map by combining risk category labels, wherein the training in the prediction process adopts a multi-objective loss function, the multi-objective training function comprises a mean square error loss term weighted by a significant region, a structure edge preservation regular term and a structure continuity consistency regular term, and the structure continuity consistency regular term is obtained by enabling the gradient directions of the final image and the stress prediction distribution map to be consistent;
Collecting a fiber structure diagram, extracting a local main structure directional diagram from the fiber structure diagram, extracting a gradient directional diagram from a stress prediction distribution diagram, calculating an angle residual diagram between the local main structure direction and the gradient directional diagram, and generating a deviation distribution index by combining the angle residual diagram and regional weights, wherein the regional weights are obtained through a space anomaly significance diagram;
And calculating an average residual value of the region from the angle residual map, taking the average residual value as a weighting coefficient and a deviation distribution index generation speed and angular speed adjusting factor, and finally generating a next period mould executing instruction based on an adjusting strategy.
The method specifically comprises the following steps:
step one, synchronously collecting demolding visual sequence and mold parameters
The step aims to construct a dynamic behavior data acquisition system for a paper support demolding process, and the acquisition content comprises a high-speed image sequenceAnd synchronously corresponding mold motion parameters(Line speed)(Angle of disengagement). The data must have extremely high time consistency and process correspondence to support subsequent behavior analysis, stress modeling, and other tasks. The core work of the step is to establish a hardware-level synchronous connection mechanism between the high-speed image acquisition equipment and the mould control system, so as to ensure that visual data and motion parameters are aligned strictly at each sampling time point.
In order to meet the characteristics of the paper support demolding process, the system adopts a lateral view angle to collect images and captures information such as displacement, deformation, rotation and the like in the demolding process. The industrial camera is Basler acA-155 um (or equivalent model) and supports external triggering, and the maximum frame rate can reach 500fps. When the camera is installed, the camera is fixed on the vertical side face of the moving direction of the die through the rigid support, the distance from the camera to the center line of the die is about 25cm, the field of view of the lens is required to cover the whole paper supporting die-out channel, and the problem of movement cutting is avoided. The camera lens uses a C-interface 12mm fixed focus lens, and an aperture is set to be f/5.6 to ensure that the depth of field is enough to capture the outline details of the paper support, and a continuous illumination structure is adopted, and an annular LED lamp array is arranged around the camera to ensure that the illumination of a shooting area is uniform and shadowless.
The image acquisition adopts an external trigger mode to control, and a TTL level pulse system with the frequency of 500Hz is built and is used as a synchronous signal source and is respectively connected to a camera trigger interface and a PLC input module. Taking the rising edge of TTL as a sampling trigger point, driving immediately after each signal trigger:
the industrial camera collects the current frame image ;
The PLC system records the die speed sampled by the current laser ranging sensor(Unit: mm/s), and reads the current die rotation angleThe angle is fed back in real time by a multi-turn encoder arranged in a servo motor, and the precision is not lower than 0.01 degrees.
For example, a certain demolding operation for 1.6 seconds, i.e. 1600ms, will generate 800 frames of images according to a 500Hz sampling frequency. The system will bind a record for each frame acquisition instant containing the current image and the speed and angle data for that instant. Taking the 123 th frame as an example, the acquisition record is:
The image acquired at the 123 st moment of 2ms is shot that the upper edge of the paper holder just leaves the mold wall;
The stretching speed of the die is 87.5 mm/s;
the included angle between the mould and the paper holder is 15.8 degrees.
All data will be written in real time to the local cache and then packaged as a structured behavioral data sequence:
;
Wherein:
Is the first RGB images of the frames, acquired by an industrial camera;
the linear speed of the die at the moment is read by a laser displacement sensor, and the unit is mm/s;
The current angle of the die is output by a servo motor encoder, and the unit is an angle (°);
The total frame number in the whole demolding process is determined by the sampling frequency and time, and is generally 600-1000 frames.
The acquisition program runs in a multithreading architecture, one thread is responsible for image caching (writing into an NVMe SSD to avoid bandwidth bottleneck), the other thread is responsible for real-time action parameter sampling and alignment, and the main control program synchronously stores with a frame number as an index.
The final output is a complete behavioral data sequence:
;
the structure is the original input for modeling of the abnormal behavior of the mold release in the subsequent step.
Step two, abnormal behavior detection and dynamic region extraction
The aim of this step is to remove the paper support by sequencing the images before the end of the demolding processAnd mold state parametersAndIn (3) the abnormal demolding behaviors possibly occurring are detected in real time, and the high-risk areas are accurately marked. These regions are directly input into the next structural stress estimation, and are important bridging points for the whole system to form a 'prediction-verification-control' closed loop. Unlike traditional visual anomaly detection, the innovation of this step is to identify whether anomalies exist or not, and to explain why the anomalies occur (dynamic behavior features), where they occur (spatial regions), how much they are (response intensity), which is particularly important for paper-backed low-contrast, complex-edge workpieces. Therefore, a combined classification and positioning network integrating the image sequence optical flow, the mold parameters and the residual difference constant is designed, the characteristics of the image change trend and the industrial process fluctuation are combined, the behavior labels can be output, and the space mapping graph can be generated to be used as the direct input of the subsequent structure prediction.
Data structure with input as step one output:
Is the firstFrame demolding process image, which is collected by a 500fps high-speed industrial camera, with the resolution of 1280×720;
the linear speed of the die is acquired by a laser displacement sensor and is in mm/s;
Is the mold separation angle, provided by the encoder, in degrees;
the total number of frames is about 600 to 1000 frames.
The data are packaged in a time alignment structure, and the images of each frame are strictly in one-to-one correspondence with the states of the mold, so that complete dynamic description is provided.
First, pairs of adjacent frame imagesIn an input TV-L1 optical flow estimation network, an inter-frame motion vector field is calculatedAnd is used as an image change trend feature. Optical flow calculation is performed on an OpenCV TV-L1 module, the window size is 17, the relaxation coefficient is 0.1, the iteration number is set to 30, and a pixel-level two-dimensional displacement map of each frame is generated and used for capturing a rapid deformation region in the demolding process.
Then, a classification network is constructed which is input as a multimodal sequence. The network comprises the following three modules:
visual characteristic extraction module for processing 16-frame image blocks by adopting 3-layer time sequence 3D convolution network Extracting space-time convolution characteristics;
optical flow coding module for corresponding optical flow sequence 2D convolution and bidirectional LSTM processing are carried out, and motion trend characteristics are extracted;
parameter branch coding module for coding state vector of mould Spliced intoSending the dynamic change trend of the state of the double-layer GRU into the double-layer GRU.
After the output of the three modules are spliced, the final characterization vector is generated through fusion of a shared attention mechanismOutputting risk category after entering classification headAnd the final layer convolution activation graph is reservedFor subsequent saliency location.
The core classification process is as follows:
;
Wherein, the As a trainable full connection layer parameter,And (5) coding vectors for the behavior characteristics after the three-branch characteristics are fused.
An innovative abnormal residual error regular term is introduced during model training, and is used for suppressing false positive samples with normal actions and abnormal results. The regularization term depends on a reference behavior template previously constructed by statistical demoulded samplesThe average code is extracted from the normal demoulding sequence to generate the product, which represents the ideal demoulding behavior. The false abnormality and the real demoulding risk are effectively distinguished by calculating the L2 residual error between the current behavior characteristic and the reference template and adding the L2 residual error into the classification loss.
The modified loss function is:
;
Wherein:
Is a standard cross entropy loss;
the average normal behavior characteristic template;
for the weight coefficient, 0.05 was taken in the experiment.
The introduction of residual regularization term makes the model judge not only the current behavior 'look abnormal', but also whether it is 'obviously deviated from the ideal demoulding process', so that the robustness is greatly improved, and especially under the scene of complex paper support shape and rapid switching, a great amount of false alarms are avoided.
For spatial interpretability, the CAM method is further used to convolve the output with a graphWeighted summation yields a saliency map:
;
Wherein:
is the last layer of convolution A plurality of channels;
the gradient weight of the channel to the final classification score is obtained by back propagation.
ResultsRepresenting the most important areas in the behavior prediction process, helps to locate the "origin of abnormal behavior", such as edge asymmetric dip in a section of the demolding trace.
The salient region map is fed into a next structure prediction model in the form of a mask for modeling local stress concentrations.
The output of the step comprises:
demolding behavior classification label The demolding is normal, mold sticking, drawing crack or collapse;
spatial saliency map Providing a spatial guiding region for the next structural modeling.
Step three, stress distribution prediction based on dynamic guidance
The step aims at classifying results according to the dynamic behaviors output in the previous stageAnd corresponding spatial high risk region mapIn combination with image frames at the end of demouldingConstructing a stress distribution prediction graph of a paper holder after demolding is completed. The method is an important intermediate link from behavior prediction to structural defect perception in the patent, and directly influences whether a potential structural instability region can be accurately judged later, so that the paper support product is promoted to be converted from 'apparent detection' to 'physical mechanics intelligent perception'.
The method is different from the traditional finite element modeling method, and the innovation of the step is to use a demolding image and abnormal demolding behavior to jointly model, and introduce a behavior significance weighting mechanism and a regional structure gradient sparse constraint term. Therefore, the accuracy of predicting the key position is improved, the high response error of the prediction graph in the normal region is greatly restrained, the 'trusted distribution' of the stress graph is realized, and the practical and available quality evaluation basis is provided for paper support production.
The input data are all from the output of step two, specifically:
The last frame image at the end of the demolding process is a sheet of size RGB images of (a);
Dynamic behavior risk classification result and value range Respectively representing a normal demolding trend, a sticking trend, a local cracking risk, a bottom collapse trend;
dynamic anomaly significance map, size and Each pixel represents a significance score for the corresponding location in the behavioral model for the same single channel thermodynamic diagram.
The step is to firstlyAnd (3) withSplicing in channel dimension to generate a four-channel input tensor. The tensor is input into a symmetrical residual-jump link structure network (based on UNet skeleton), the model includes four-stage downsampling and four-stage upsampling modules, and each stage of coding stage includes two layersThe convolution+ReLU+ BatchNorm module, the number of channels is 64/128/256/512 in turn, the decoding stage uses deconvolution reconstruction, and the jump features from the encoder are fused.
To classify labelsIntegrating the network prediction process, we design a kind of 'class attention graph mechanism'Conversion to a 3-channel spatial template map by a lookup tableThe template derives from the average stress distribution pattern for each type of defect in the training set (e.g., the "collapse risk" is typically concentrated in the bottom centerline area of the tray) and is then input as the fifth pass along with the original input tensor. The final input is a five-channel tensor:
;
The predicted output of the stress diagram is Each positionIndicating the relative stress degree, normalized toInterval.
The training process adopts a multi-objective loss function, wherein the first is a mean square error loss term weighted by a significant region, the second is a structure edge retention regularization term, and the third is a structure continuity consistency regularization term innovatively designed by usUsed for restraining the consistency of the image structure and the gradient direction of the stress diagram.
The final loss function is as follows:
;
Wherein:
Predicting a value for the current model;
Is a reference stress graph (generated by manually combining simulation data);
weighting coefficients (taking 3) for the saliency map, emphasizing the importance of the high risk region;
regular weight for edge (0.01);
is a structural consistency term coefficient (0.1);
While the structural consistency term The calculation is as follows:
;
Wherein:
Representing an image The gradient direction at that location (calculated by Sobel operator);
Representing the gradient direction of the predicted stress pattern at that location;
This term encourages stress patterns to remain consistent with image contour directions at structure boundaries, avoiding distortion problems of "predictive pattern non-stick image structures".
The regular term is one of the core innovation points of the step, and the interpretability and the stability of the prediction graph are improved by directly combining the image texture structure and the stress direction consistency.
The model adopts an Adam optimizer, and the initial learning rateTraining 30 rounds on 1000 groups of demolding images and stress reference data thereof, and finally, on a verification set, the MAE is 0.021, so that the structural matching consistency is improved to more than 84%, and the method has good generalization capability.
The output of this step is a predicted stress distribution mapEach pixel value thereof represents the potential stress concentration degree caused by abnormal demolding, and the higher the value is, the weaker the structure of the region is. The image is used as a comparison basis for verifying the structural image in the next step and is also used for judging whether the control feedback flow should be entered or not.
Step four, verifying structural image and confirming prediction precision
The main aim of the step is to predict the stress of the paper support structure generated in the last stepAnd verifying the structural image level. Compared with the simple 'result visualization' or 'pixel coincidence rate analysis' in the traditional industrial vision system, the method combines the non-uniformity of the molding material, the non-structural defect concealment and the local error volatility of the prediction model of the paper support in the demolding process, designs a multi-scale direction matching verification mechanism for structure tensor guidance, carries out space direction field consistency matching on a stress prediction graph and a paper support fiber structure diagram obtained by polarized light imaging, and creatively introduces angle tensor residual error regular based on main direction differenceAnd a bias distribution index combining the direction difference and the area weightFor quantifying and explicitly outputting the "prediction confidence level" and the "inconsistent distribution of stress field and structure diagram".
To enable the verification mechanism to fully utilize the material structure information, we first follow fromExtracting local principal structural pattern. Using structure tensor analysis method, inThe directionality of each pixel is calculated in the Gaussian weighted window, and a standard structure tensor formula is adopted:
;
Wherein:
respectively obtaining structure tensor elements by carrying out Sobel processing on the image at the position and then carrying out Gaussian kernel smoothing on the sum of squares and products of the sum of the gradients;
is a numerical stability term, preventing denominator from approaching 0;
indicating the principal direction of the local fiber at the position, the numerical range is 。
Also, fromThe direction of the medium extraction gradientAnd reflecting the stress expansion main axis direction predicted by the model. The direction is thatApplying Sobel operator to obtain gradient map, and calculating gradient direction angle。
Then calculate the angular residual map between the two directional fields:
;
The residual map is used for capturing the structural difference of the two images at the directional level, wherein the closer to 1 is the larger the deviation of the two directions is, and the closer to 0 is the direction coincidence is.
Different from the traditional IoU or pixel level difference, the angle residual error can reflect the consistency of the force transmission path in the paper holder and the actual stress direction of the material, and has stronger interpretation. Especially whenThe high stress in a certain area is predicted, and the fiber arrangement in the area is in staggered disorder and direction unclear, so that the model prediction can be judged to be false.
On the basis, we further construct a region deviation distribution index combining the directional residual image and the region saliency weight imageThe method is used for measuring the predicted deviation severity of the model in the key area in the whole image range:
;
Wherein:
Indicating whether the area is a high risk concern area in the demolding behavior or not for the dynamic saliency map generated in the step two;
is a directional residual error diagram;
Is that Normalization constant of image pixel sums (ensuringAt the position of);
The larger the model prediction is, the larger the difference between the model prediction and the real direction of the structure in a key area is, and the lower the reliability is.
The design of the index is a core creation point of the patent on visual structure verification, and the model output of the previous step and the current actual structure observation result are combined, so that whether the model is credible or not can be quantized, and the index can be used for judging whether feedback adjustment needs to be executed or not in the subsequent control step.
For example, if the region is predictedAnd (3) withAt the position ofThe strong response area has obvious deviation) The model judgment is considered unreliable, the current mould demoulding parameters are required to be kept, and if the deviation is smaller, the model judgment is considered unreliable) The parameter adaptive update module may be triggered.
The output of this step is two variables:
the direction deviation mapping diagram of each pixel point is used for regional structure deviation positioning;
The overall verification index is used for measuring whether the prediction graph is reliable in structure or not.
Generating a demoulding control strategy and outputting a control instruction
The aim of the step is to use the structure verification result output in the previous stage for the adjustment of the actual control strategy to generate the action execution instruction of the demoulding process in the next period, namely the new mould movement speedAnd angle change parameterAnd is directly executed by the industrial PLC system. The focus of this step is how to use the structural direction residualsAnd overall verification deviation indexWhile ensuring the production takt, fine and effective self-adaptive adjustment is made to the mold stripping behavior, so that slight risk accumulation is prevented from becoming a defect.
In order to achieve smooth and reliable control parameter adjustment, this step designs a residual-guided adaptive strategy update logic. The logic passes through two directions:
in a high deviation area of the structure, the speed of the die is properly reduced, and the pulling impact on the paper holder is slowed down;
And when the integral prediction reliability is not high, the angular transition amplitude of the die is slowed down, and the cracking caused by too fast structural deformation is avoided.
We construct the average residual value of the structure mismatch area as the weighting coefficient:
;
and then according to the value and the integral deviation index The combined generation of speed and angle adjustment factors:
;
Wherein:
the rate compression coefficient is controlled so that, Is the minimum allowable compression ratio;
the slowing down coefficient of the angle change amplitude is controlled, Is the minimum value;
smaller coefficients indicate greater deviations and should be handled with greater care.
Finally, the next period of die execution instruction is generated as follows:
;
The controller will make the group of As a setting target of the next demoulding process, the setting target is written into an industrial control system through a PLC instruction module (Modbus or CAN bus), and the specific movement is realized by a servo motor.
It is worth mentioning that the light-weight adjustment mechanism for adjusting and controlling the mechanical behavior based on the image verification result is used herein, and does not need reverse modeling or rely on a predefined demolding physical process, so that the method has good expansibility and robustness in industrial deployment.
The final output is the control execution parameterDirectly acting on the motion driver of the mold control system.
The technical scheme provided by the embodiment of the application has at least the following technical effects or advantages that the application provides a visual intelligent quality detection method and a visual intelligent quality detection system for a paper holder demolding process. The system analyzes whether abnormal trends exist in demolding behaviors in real time by collecting high-frequency image sequences of the paper holder in the demolding process and combining action parameter information of the mold, and predicts potential quality problems possibly caused by the behaviors. At the same time, the application further evaluates the stress variations and possible deformation areas of the tray at different stages from the visual data, thereby identifying those structural defects that are about to be or are forming but are not yet noticeable to the naked eye. In addition, the system can also feed back the detection result to the demolding control unit to realize intelligent adjustment of key parameters such as mold speed, separation angle and the like, so that a closed loop quality control chain of prediction, detection, verification and regulation is formed. By combining process modeling and structure sensing, the application effectively makes up the defects of the traditional paper support defect detection means based on static images in terms of time lag, structure invisibility, control unreachability and the like, and remarkably improves the intelligent level and detection coverage rate of paper support quality detection.
The embodiment of the application also provides a paper support production quality detection system based on visual processing, as shown in fig. 2, comprising:
the data acquisition module acquires original data, wherein the original data comprises an image sequence and corresponding mould motion parameters, aligns the image sequence with the motion parameters at a time point and outputs a behavior data sequence;
the anomaly detection module extracts space-time convolution characteristics, motion trend characteristics and state change trends from the behavior data sequence, splices the space-time convolution characteristics, the motion trend characteristics and the state change trends to obtain a final characterization vector, classifies risk categories of the final characterization vector, outputs risk category labels of the final characterization vector, and simultaneously performs weighted summation on a final layer of convolution activation diagram of the final characterization vector to obtain a space anomaly significance diagram;
The stress prediction module is used for splicing a final image of the demolding process with the spatial abnormality significance map and constructing a stress prediction distribution map by combining a risk category label, wherein the training of the prediction process adopts a multi-objective loss function, the multi-objective training function comprises a mean square error loss term weighted by a significant region, a structure edge preservation regular term and a structure continuity consistency regular term, and the structure continuity consistency regular term is obtained by enabling the gradient directions of the final image and the stress prediction distribution map to be consistent;
the image verification module acquires a fiber structure diagram, extracts a local main structure direction diagram from the fiber structure diagram, extracts a gradient direction diagram from a stress prediction distribution diagram, calculates an angle residual diagram between the local main structure direction and the gradient direction diagram, and combines the angle residual diagram and regional weight to generate a deviation distribution index;
And the strategy generation module is used for calculating an area average residual value from the angle residual map, taking the area average residual value as a weighting coefficient and a deviation distribution index generation speed and angular speed regulating factor, and finally generating a next period mould execution instruction based on an adjustment strategy.
The system analyzes whether abnormal trends exist in demolding behaviors in real time by collecting high-frequency image sequences of the paper holder in the demolding process and combining action parameter information of the mold, and predicts potential quality problems possibly caused by the behaviors. At the same time, the invention further evaluates the stress variations and possible deformation areas of the tray at different stages from the visual data, thereby identifying those structural defects that are about to be or are forming but are not yet noticeable to the naked eye. In addition, the system can also feed back the detection result to the demolding control unit to realize intelligent adjustment of key parameters such as mold speed, separation angle and the like, so that a closed loop quality control chain of prediction, detection, verification and regulation is formed.
The foregoing is only a preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art, who is within the scope of the present invention, should make equivalent substitutions or modifications according to the technical scheme of the present invention and the inventive concept thereof, and should be covered by the scope of the present invention.

Claims (8)

1.基于视觉处理的纸托生产质量检测方法，其特征在于，包括：1. A method for quality inspection of paper tray production based on vision processing, characterized in that it includes: 采集原始数据，所述原始数据包括图像序列及对应的模具运动参数；将图像序列与运动参数以时间点对齐，输出行为数据序列；Collect raw data, which includes image sequences and corresponding mold motion parameters; align the image sequences and motion parameters by time points to output a behavioral data sequence. 从行为数据序列中提取时空卷积特征、运动趋势特征和状态变化趋势，并拼接得到最终表征向量，对最终表征向量进行风险类别分类，并输出其风险类别标签，同时对最终表征向量的最后一层卷积激活图加权求和得到空间异常显著性图；Spatiotemporal convolutional features, motion trend features, and state change trends are extracted from behavioral data sequences and concatenated to obtain the final representation vector. The final representation vector is classified into risk categories and its risk category label is output. At the same time, the spatial anomaly saliency map is obtained by weighted summation of the last layer of convolutional activation map of the final representation vector. 将脱模过程的最终图像与空间异常显著性图拼接，并结合风险类别标签构建应力预测分布图，其中，预测过程的训练采用多目标损失函数，所述多目标损失函数包括以显著区域加权的均方差损失项、结构边缘保留正则项和结构连续性一致性正则项，所述结构连续性一致性正则项通过使最终图像和应力预测分布图的梯度方向一致得到；The final image of the demolding process is stitched together with the spatial anomaly saliency map, and a stress prediction distribution map is constructed by combining the risk category label. The training of the prediction process adopts a multi-objective loss function, which includes a mean squared error loss term weighted by salient regions, a structural edge preservation regularization term, and a structural continuity consistency regularization term. The structural continuity consistency regularization term is obtained by making the gradient directions of the final image and the stress prediction distribution map consistent. 采集得到纤维结构图，从纤维结构图中提取得到局部主结构方向图，从应力预测分布图中提取梯度方向图，然后计算局部主结构方向和梯度方向图之间的角度残差图，结合角度残差图和区域权重生成偏差分布指数；所述区域权重通过空间异常显著性图得到；The fiber structure map is collected, the local main structure orientation map is extracted from the fiber structure map, and the gradient orientation map is extracted from the stress prediction distribution map. Then, the angle residual map between the local main structure orientation map and the gradient orientation map is calculated. The deviation distribution index is generated by combining the angle residual map and the regional weights. The regional weights are obtained through the spatial anomaly significance map. 对角度残差图计算得到区域平均残差值，并将其作为加权系数与偏差分布指数生成速度和角速度调节因子，并基于调整策略最终生成下一周期模具执行指令。The average residual value of the region is calculated from the angular residual map and used as a weighting coefficient and a deviation distribution index to generate the speed and angular velocity adjustment factor. Based on the adjustment strategy, the mold execution command for the next cycle is finally generated. 2.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述模具运动参数包括模具运动的线速度以及模具与纸托之间的脱离角度。2. The paper tray production quality inspection method based on vision processing according to claim 1, wherein the mold motion parameters include the linear velocity of the mold motion and the separation angle between the mold and the paper tray. 3.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述运动趋势特征通过对对应的光流序列进行处理得到，所述光流序列为帧间运动向量场组合得到，所述帧间运动向量场为将行为数据序列中相邻帧图像输入光流估计网络中计算得到。3. The paper tray production quality inspection method based on visual processing according to claim 1, characterized in that the motion trend feature is obtained by processing the corresponding optical flow sequence, the optical flow sequence is obtained by combining inter-frame motion vector fields, and the inter-frame motion vector field is calculated by inputting adjacent frame images in the behavioral data sequence into an optical flow estimation network. 4.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述风险类别标签包括正常、粘膜、拉裂和塌陷。4. The paper tray production quality inspection method based on vision processing according to claim 1, wherein the risk category label includes normal, adhesive, tear, and collapse. 5.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述风险类别分类训练过程中引入异常残差正则项，所述异常残差正则项通过一个事先基于统计脱模样本构建的参考行为模板，由正常脱模序列提取平均编码后生成；计算当前行为特征与参考行为模板的L2残差并加入到分类损失中，以区分伪异常与真实脱模风险。5. The paper tray production quality inspection method based on visual processing according to claim 1, characterized in that, during the risk category classification training process, an abnormal residual regularization term is introduced, which is generated by extracting the average code of the normal demolding sequence through a reference behavior template pre-constructed based on a statistical demolding template; the L2 residual between the current behavior feature and the reference behavior template is calculated and added to the classification loss to distinguish between pseudo-anomalies and real demolding risks. 6.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述局部主结构方向图通过结构张量分析方法，在高斯加权窗口内计算每个像素的方向性得到。6. The paper tray production quality inspection method based on vision processing according to claim 1, characterized in that the local master structure orientation map is obtained by calculating the orientation of each pixel within a Gaussian weighted window using the structural tensor analysis method. 7.根据权利要求1所述的基于视觉处理的纸托生产质量检测方法，其特征在于，所述调整策略包括：7. The paper tray production quality inspection method based on vision processing according to claim 1, characterized in that the adjustment strategy includes: 对于结构高偏差区域，降低模具速度，减缓对纸托的拉拽冲击；For areas with high structural deviations, reduce the mold speed to mitigate the pulling and impact on the paper tray; 对于整体预测可信度不高时，减缓模具角度转变幅度，避免结构形变过快造成拉裂。When the overall prediction reliability is not high, the change in mold angle should be slowed down to avoid excessive structural deformation that could cause tearing. 8.基于视觉处理的纸托生产质量检测系统，其特征在于，包括：8. A paper tray production quality inspection system based on vision processing, characterized in that it includes: 数据采集模块，所述数据采集模块采集原始数据，所述原始数据包括图像序列及对应的模具运动参数；将图像序列与运动参数以时间点对齐，输出行为数据序列；The data acquisition module acquires raw data, including image sequences and corresponding mold motion parameters; it aligns the image sequences and motion parameters by time points and outputs a behavioral data sequence. 异常检测模块，所述异常检测模块从行为数据序列中提取时空卷积特征、运动趋势特征和状态变化趋势，并拼接得到最终表征向量，对最终表征向量进行风险类别分类，并输出其风险类别标签，同时对最终表征向量的最后一层卷积激活图加权求和得到空间异常显著性图；An anomaly detection module extracts spatiotemporal convolutional features, motion trend features, and state change trends from behavioral data sequences, and concatenates them to obtain a final representation vector. The final representation vector is classified into risk categories, and its risk category label is output. At the same time, the last layer of convolutional activation map of the final representation vector is weighted and summed to obtain a spatial anomaly saliency map. 应力预测模块，所述应力预测模块将脱模过程的最终图像与空间异常显著性图拼接，并结合风险类别标签构建应力预测分布图，其中，预测过程的训练采用多目标损失函数，所述多目标损失函数包括以显著区域加权的均方差损失项、结构边缘保留正则项和结构连续性一致性正则项，所述结构连续性一致性正则项通过使最终图像和应力预测分布图的梯度方向一致得到；The stress prediction module stitches the final image of the demolding process with the spatial anomaly saliency map and constructs a stress prediction distribution map by combining it with risk category labels. The training of the prediction process adopts a multi-objective loss function, which includes a mean squared error loss term weighted by salient regions, a structural edge preservation regularization term, and a structural continuity consistency regularization term. The structural continuity consistency regularization term is obtained by making the gradient directions of the final image and the stress prediction distribution map consistent. 图像验证模块，所述图像验证模块采集得到纤维结构图，从纤维结构图中提取得到局部主结构方向图，从应力预测分布图中提取梯度方向图，然后计算局部主结构方向和梯度方向图之间的角度残差图，结合角度残差图和区域权重生成偏差分布指数；所述区域权重通过空间异常显著性图得到；The image verification module acquires a fiber structure image, extracts a local principal structure orientation pattern from the fiber structure image, extracts a gradient orientation pattern from the stress prediction distribution image, calculates the angular residual image between the local principal structure orientation pattern and the gradient orientation pattern, and generates a deviation distribution index by combining the angular residual image and the region weights; the region weights are obtained through a spatial anomaly significance map. 策略生成模块，所述策略生成模块对角度残差图计算得到区域平均残差值，并将其作为加权系数与偏差分布指数生成速度和角速度调节因子，并基于调整策略最终生成下一周期模具执行指令。The strategy generation module calculates the regional average residual value from the angular residual map and uses it as a weighting coefficient and deviation distribution index to generate the speed and angular velocity adjustment factor. Based on the adjustment strategy, it finally generates the mold execution command for the next cycle.
CN202510710915.XA 2025-05-29 2025-05-29 A Vision-Based Method and System for Quality Inspection of Paper Tray Production Active CN120496055B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202510710915.XA CN120496055B (en) 2025-05-29 2025-05-29 A Vision-Based Method and System for Quality Inspection of Paper Tray Production

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202510710915.XA CN120496055B (en) 2025-05-29 2025-05-29 A Vision-Based Method and System for Quality Inspection of Paper Tray Production

Publications (2)

Publication Number Publication Date
CN120496055A CN120496055A (en) 2025-08-15
CN120496055B true CN120496055B (en) 2026-01-30

Family

ID=96667307

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202510710915.XA Active CN120496055B (en) 2025-05-29 2025-05-29 A Vision-Based Method and System for Quality Inspection of Paper Tray Production

Country Status (1)

Country Link
CN (1) CN120496055B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120746403B (en) * 2025-09-03 2025-12-02 江苏数兑科技有限公司 Large model driven data management decision support system
CN120890988B (en) * 2025-09-29 2025-12-09 国网甘肃省电力公司兰州供电公司 Intelligent Substation Inspection Robot and Fault Handling System Based on Multimodal Perception

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112557406A (en) * 2021-02-19 2021-03-26 浙江大胜达包装股份有限公司 Intelligent inspection method and system for paper product production quality
CN119205719A (en) * 2024-10-25 2024-12-27 河南铭视科技股份有限公司 Intelligent machine vision detection method, system and storage medium based on image processing

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114022442B (en) * 2021-11-03 2022-11-29 武汉智目智能技术合伙企业(有限合伙) Unsupervised learning-based fabric defect detection algorithm

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112557406A (en) * 2021-02-19 2021-03-26 浙江大胜达包装股份有限公司 Intelligent inspection method and system for paper product production quality
CN119205719A (en) * 2024-10-25 2024-12-27 河南铭视科技股份有限公司 Intelligent machine vision detection method, system and storage medium based on image processing

Also Published As

Publication number Publication date
CN120496055A (en) 2025-08-15

Similar Documents

Publication Publication Date Title
CN120496055B (en) A Vision-Based Method and System for Quality Inspection of Paper Tray Production
CN109840900B (en) A fault online detection system and detection method applied to intelligent manufacturing workshops
US12350888B2 (en) AI-powered autonomous 3D printer
CN120833319B (en) Mobile phone silica gel shell appearance detection system based on optical image sensor
CN120587085B (en) A Smart Dispensing Control Method Based on Adaptive Trajectory and Dynamic Fault-Tolerant Control
CN109693140A (en) A kind of intelligent flexible production line and its working method
CN120726054B (en) Surface anomaly detection method based on multi-mode dynamic fusion and two-way collaborative detection
CN117054446A (en) Comprehensive detection device for photovoltaic cell screen plate
CN118808079A (en) A dispensing control system for intelligent conversion of visual templates and processing trajectories
CN110111332A (en) Collagent casing for sausages defects detection model, detection method and system based on depth convolutional neural networks
CN117830292A (en) Air conditioner indoor unit defect detection system and control method and storage medium thereof
WO2025256668A2 (en) An artificial intelligence (ai) -enabled real-time defect detection system for additive manufacturing based on multimodal data and the method thereof
CN120080368B (en) PVC rigid sheet high-speed slitting quality monitoring system based on AI visual inspection
CN120339701B (en) A surface mount inspection system for semiconductor packaging
CN119359630A (en) An intelligent detection method and system for steel surface defect identification
CN116843615A (en) Lead frame intelligent total inspection method based on flexible light path
CN117301631A (en) Color box printing automatic folding intelligent management system
Zhang Dual-channel capsule shape detection system based on computer vision and optical imaging technologies
CN120932026B (en) Automatic identification method, storage medium and program product for flying steel of rod and wire rod in rolling mill workshop
CN120765551B (en) Image Detection-Based Method and System for Detecting Double-Outlet Zipper Bags
CN120689699B (en) An AI visual inspection method based on supervised and unsupervised deep learning
CN120741483B (en) Intelligent detection method and system for conveyor line based on machine vision
CN119260738A (en) 3D visual recognition method for industrial robots based on deep learning
CN120669503A (en) Method for measuring rise and fall time of electric railing machine based on neural network model
CN120744808A (en) Industrial production detection method and system based on multi-mode model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant