CN107274401B - High-resolution SAR image ship detection method based on visual attention mechanism - Google Patents

Abstract

The invention discloses a high-resolution SAR image ship detection method based on a visual attention mechanism, which belongs to the field of radar image target detection. Mainly aiming at the fact that the existing high-resolution SAR image target detection cannot meet the intelligent requirements and real-time requirements of detection, the Fourier transform based on the frequency domain extracts the spectral residual part of the image containing the target information to extract the visual saliency model; The saliency map is binarized and the region of interest is extracted. From the perspective of classification, a local maximum a posteriori probability classifier is designed to detect the potential target region, and the detection is realized by parameter estimation and decision criteria. This method can improve the real-time and accuracy of ship target detection in high-resolution SAR images, and effectively avoid high false alarm problems in detection.

Description

A high-resolution SAR image ship detection method based on visual attention mechanism

technical field

The invention belongs to the field of synthetic aperture radar image target detection, relates to a detection method that meets the real-time detection requirements of radar image targets, and is suitable for ship target detection and monitoring in complex sea conditions including a large number of uneven sea clutter areas and speckle noise. high-resolution synthetic aperture radar imagery.

Background technique

Synthetic aperture radar (SAR) has the characteristics of all-day, all-weather, and large-scale, and is an important part of ocean monitoring and surveillance, among which ship target detection has increasingly become a research hotspot. SAR image ship target detection is the premise and foundation of its classification and recognition, and is an important aspect of SAR image application.

With the launch and operation of new-generation SAR sensors such as Radarsat-2, TerraSAR-X and Gaofen-3, SAR is gradually developing towards high resolution, large width, and multi-polarization. However, as the size of high-resolution SAR images gradually increases, the traditional image-based point-by-point calculation is slow, and the contradiction between the large amount of SAR image information and the limited computer processing capability makes it difficult to meet the requirements of real-time processing. . Secondly, the detection accuracy of traditional low-resolution SAR image detection methods is not high when applied to high-resolution SAR images at this stage, and it still exists in improving the false alarms caused by speckle noise and uneven sea clutter background on the detection results. Insufficient, it is difficult to meet the needs of accurate and intelligent detection of targets in images.

SUMMARY OF THE INVENTION

The invention proposes a high-resolution SAR image ship detection method based on a visual attention mechanism, which aims to solve the problem of multiple false alarms in the existing SAR image target detection, and meet the requirements of detection intelligence and real-time performance.

The high-resolution SAR image ship detection method based on the visual attention mechanism described in the present invention specifically includes the following technical measures: the acquisition of the visual saliency model adopts a global saliency area calculation algorithm of the spectral residual method, based on the frequency The fast Fourier transform of the domain is realized, and the part of the spectrum of an image that is different from the same shape, that is, the spectral residual part containing the target information, is extracted; The first step is the binarization of the saliency map and the extraction of the region of interest. To complete this step, it is necessary to perform two threshold segmentation processing on the saliency map. First, the potential ship region in the visual saliency map is segmented through the first threshold value, and the second threshold value segmentation uses two thresholds to classify the pixels in the saliency map. , by screening the classified parts to accurately complete the approximate fitting of the gray histogram of the subsequent target area and background area; according to the theory of machine learning, analyze the ship target detection problem from the perspective of classification, and design a local The maximum posterior probability classifier further performs ship target detection on salient areas in the image, and transforms the target detection problem into a binary hypothesis testing problem for each pixel in the potential target area; The conditional probability of the measured pixels and the prior probability of the pixels to be measured belong to each category are estimated and calculated, and combined with the estimated classifier parameters, the pixels in the potential target area of the ship in the significant area are detected twice.

A high-resolution SAR image ship detection method based on a visual attention mechanism proposed by the present invention can improve the real-time performance and accuracy of high-resolution SAR image ship target detection, and use visual attention theory and machine learning theory for image processing. , which can effectively avoid the high false alarm problem in detection.

Description of drawings

Figure 1 is a flow chart of a high-resolution SAR image ship detection based on visual attention mechanism;

Figure 2 is a schematic diagram of a high-resolution SAR image ship detection based on a visual attention mechanism.

Detailed ways

The technical solution of a high-resolution SAR image ship detection method based on a visual attention mechanism proposed by the present invention includes the following steps:

Step 1.1: The calculation model uses the global search method for the saliency region extraction algorithm, which is a visual saliency region extraction method based on the image frequency domain, which is realized by fast Fourier transform. Different from the part of the same shape, which is the spectral residual part containing the target information, the visual saliency model will extract this residual part; assuming that I(x) is an image, the spectral FFT[I(x)] of the image can be decomposed into two parts, the amplitude spectrum A(f) and the phase spectrum P(f);

P(f)=FFT[I(x)]/|{FFT[I(x)]}| (1)

R(f)=BP(f)·P(f) (2)

S(f)=FFT ^-1 [R(f)] (3)

Among them, FFT and FFT ^-1 represent the fast Fourier transform of the image and its inverse transform, P(f) is the phase spectrum of the original image, R(f) is the spectral residual, BP(f) is the band-pass filter, and the model A Gaussian filter with a center frequency of f ₀ and a cut-off frequency of Δf is used in the present invention; S(f) is the saliency map; the spectral residual saliency calculation in the ship detection method based on the visual attention mechanism in the high-resolution SAR image in the present invention The model mainly includes two steps of operation: one is the normalization processing of the original spectrum, and the other is the bandpass filtering in the frequency domain;

Step 1.2: Use the spectral residual method in the previous operation to obtain the visual saliency map, and perform two-step operations on the obtained saliency map: one is the binarization of the saliency map; the other is the extraction of the region of interest. We use two threshold segmentations, firstly segment the salient regions in the visual saliency map, so as to screen out the salient regions in the image from the visual attention computational model, which are potential ship regions. The second threshold segmentation divides the pixels in the saliency map by setting two thresholds, and more accurately completes the approximation of the gray histogram of the subsequent target area and background area.

Step 2.1: Analyze the ship target detection problem from the idea of classification in machine learning. Ship target detection is a two-class classification problem. Design a local maximum posterior probability classifier to threshold the potential ships in the visually salient region. Secondary detection of target area: According to Bayesian theory, the problem of ship target detection in high-resolution SAR images is transformed into a binary hypothesis test problem for the data vector x; the data samples are divided into two categories, and the sample categories ω _i are ω ₁ and ω ₀ , the sample category marked as target is ω ₁ , and the sample category marked as background is ω ₀ ; let P(ω _i ) represent the prior probability that the input pixel belongs to ω _i Yeas' criterion is:

Among them, P(ω ₁ |x) and P(ω ₀ |x) refer to the posterior probability that the detected pixel is the target and the background, respectively, P(x|ω _i ) is the conditional probability under the given category ω _i , P(x) refers to the probability of obtaining a pixel;

According to the Bayesian criterion and the maximum posterior probability estimation criterion, the classifier is defined as:

The conditions for the existence of the target are:

Conversely, the condition satisfied when the target does not exist is:

The decision criterion of the local maximum posterior probability classifier is:

Step 2.2: Find the conditional probability P(x|ω _i ) of the pixel to be measured under a given category:

To estimate the parameters of the local classifier, the first step is to obtain the conditional probability P(x|ω _i ), which is the probability density function of the target and the background, and set two thresholds T ₁ , T ₂ , where T ₂ >T ₁ ; Perform threshold segmentation on the saliency map and fit the gray histogram of the approximate target and background area in the corresponding original image. This process models both the background and the target:

a) Use two thresholds to segment the saliency map, extract all pixels in the corresponding saliency map greater _than the threshold T1 in the image to be detected, and obtain a grayscale histogram, which is carried out by using distribution models such as Gamma distribution, Weibull distribution, and Log-Normal distribution. Approximate grayscale histogram fitting to obtain the probability distribution of ship targets;

b) As in the method of sub-step a) above, extract all the pixels in the area smaller than the threshold T ₂ in the corresponding saliency map in the original image as an approximation of the background area, and perform gray histogram fitting as an approximation of the background probability distribution;

c) Obtain the probability density function of the ship target and the clutter background according to the probability distribution obtained by the histogram fitting, and obtain the conditional probability P(x|ω _i ) of the target and the background that needs to be obtained in the classifier.

Step 2.3: Obtain the prior probability P(ω _i ) that the pixel to be measured belongs to each category:

The prior probability that the pixel to be tested belongs to ω _i is obtained through a sliding window, and the prior probability P(ω _i ) is defined as:

Among them, x _t represents the current gray value of the pixel to be detected, and x ₁ , x ₂ ,...,x _N are all the pixels belonging to the background area in the sliding window. G is the maximum gray level of the SAR image, a is an empirical parameter for adjusting the prior probability, a ∈ (0,1]; in this step, a sliding window is used to calculate the prior probability of the target potential area, and the sliding window design Each parameter in the SAR image needs to be determined according to the pixel area, size and distribution of the ship target in the SAR image;

Step 2.4: Combine the obtained prior probability and conditional probability, realize the secondary detection of the pixels in the potential target area of the ship in the salient area, and use the designed local maximum a posteriori probability classifier to perform all pixel points in the salient area in the image. Ship target detection, get the final detection result.

Claims (2)

1. A high-resolution SAR image ship detection method based on a visual attention mechanism is characterized by comprising the following steps:

step 1: designing a global search high-resolution SAR image salient region detection algorithm, obtaining a spectrum residual visual salient calculation model through spectrum normalization processing and frequency domain band-pass filtering, and quickly obtaining a visual interesting region;

step 2: combining a binary hypothesis testing thought in Bayes theory, designing a local maximum posterior probability classifier, and completing pixel two classification in a significant region through parameter estimation and decision criteria to realize target detection;

the step 2 specifically comprises the following substeps:

step 2.1: analyzing the ship target detection problem from the classification angle, designing a local maximum posterior probability classifier to further detect the ship target in the salient region in the image: converting a target detection problem into a binary hypothesis test problem of a data vector x according to a Bayesian theory; the data samples are divided into two classes, and the sample class is omega respectively₁And ω₀Let P (ω)_i) Indicating that the input pixel belongs to ω_iThe bayesian criterion of this binary hypothesis detection is:

wherein, P (ω)₁| x) and P (ω)₀| x) refers to the posterior probability of the detected pixel being the target and background, respectively, P (x | ω)_i) Is in a given category ω₀Conditional probability of, p (x) refers to the probability of acquiring a pixel;

according to the Bayesian criterion and the maximum posterior probability criterion, the classifier can be defined as:

the conditions met when the target is present are:

the decision criterion adopted by the maximum a posteriori probability classifier is as follows:

step 2.2: solving the conditional probability P (x | omega) of the pixel point to be detected under the given category_i)：

Conditional probability P (x | ω |)_i) Namely, setting two threshold values T as probability density functions of the target and the background₁、T₂Wherein T is₂>T₁；

a) The two thresholds are adopted to segment the saliency map, and the saliency map more than the threshold T in the corresponding saliency map in the image to be detected is extracted₁Obtaining a gray level histogram of all partial pixels; fitting approximate gray level histogram of Gamma distribution, Weibull distribution and Log-Normal distribution to obtain probability distribution of the ship target;

b) extracting the corresponding saliency map in the original map smaller than the threshold T in the same way as the sub-step a) above₂Performing gray level histogram fitting on all pixels in the region to serve as approximation of background probability distribution;

c) solving a probability density function of the ship target and the clutter background according to the probability distribution obtained by the histogram fitting;

step 2.3: solving prior probability P (omega) that pixel points to be detected belong to each category_i)：

Obtaining omega of pixel point to be detected through a sliding window_iA priori probability of (a), a priori probability P (ω)_i) Is defined as:

wherein x is_tRepresenting the gray value, x, of the current pixel to be detected₁、x₂,...,x_NIs all pixel points belonging to the background area in the sliding window, G is the maximum value of the gray level of the SAR image, a is an empirical parameter for adjusting the prior probability, a ∈ (0, 1)](ii) a At this stepIn the method, a sliding window is adopted to calculate the prior probability of a target potential region, and each parameter in the sliding window design needs to be determined according to the pixel area, the size and the distribution condition of a ship target in an SAR image;

step 2.4: and (3) combining the obtained prior probability and conditional probability to realize secondary detection on the pixel points in the ship potential target area of the salient area, carrying out ship target detection on all the pixel points in the salient area in the image by adopting a designed local maximum posterior probability classifier in the detection, and obtaining a final detection result through a judgment criterion.

2. The visual attention mechanism-based high-resolution SAR image ship detection method according to claim 1, wherein the step 1 specifically comprises the following substeps:

step 1.1: the model uses a salient region calculation algorithm of global range search, the calculation model is based on frequency domain processing and is realized by adopting fast Fourier transform, and the part which is different from the part with the same shape in the frequency spectrum of one image, namely the frequency spectrum residual part containing target information is extracted; assuming that I (x) is an image, the frequency spectrum FFT [ I (x) ] of the image is decomposed into two parts of an amplitude spectrum A (f) and a phase spectrum P (f),

P(f)＝FFT[I(x)]/|{FFT[I(x)]}| (1)

R(f)＝BP(f)·P(f) (2)

S(f)＝FFT^-1[R(f)](3)

wherein, FFT and FFT^-1Representing the fast Fourier transform and the inverse transform of the image, P (f) is the phase spectrum of the original image, R (f) represents the residual of the frequency spectrum, BP (f) is a band-pass filter, and a center frequency f is adopted in the model₀A Gaussian filter with cut-off frequency delta f, S (f) is a saliency map, and the spectrum residual saliency model mainly comprises two steps of operations: normalization processing of frequency spectrum and frequency domain band-pass filtering;

step 1.2: obtaining a visual saliency map by a previous step of spectrum residual error method, and carrying out two steps of operations on the obtained saliency map: firstly, extracting and segmenting a salient region in a visual salient image, and screening out a potential ship target region in the image according to a visual attention calculation model; the second threshold segmentation is realized by two thresholds, pixels in the saliency map are classified, and the gray histogram approximate fitting of a subsequent target region and a background region is completed more accurately.

Images