Non-cooperative vehicle target detection method in SAR images based on cooperative saliency

Abstract

The present invention relates to the technical field of image data processing and discloses a method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency. The method uses auxiliary data and test data to calculate collaborative saliency to obtain salient features of non-cooperative targets and complete detection. The method applies collaborative saliency to the task of target detection in SAR images, calculates image contrast features and joint features to obtain a preliminary collaborative saliency map, and finally uses weights generated from the collaborative saliency map of auxiliary data to further suppress background and enhance targets, thereby improving detection performance. The method provides a new solution to the problem of non-cooperative target detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and more particularly to a method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency. Background Art

[0002] Synthetic Aperture Radar (SAR), with its unique advantages such as long detection range, high resolution, and all-day, all-weather coverage, has become a widely used remote sensing information acquisition technology, possessing significant value in both military reconnaissance and civilian applications. Vehicle target detection is a key technology in SAR automatic target recognition (SAR ATR) research, providing valuable support for applications such as battlefield reconnaissance and surveillance, urban traffic monitoring, and emergency rescue and search. Therefore, research on vehicle target detection in complex scenes is of great significance. In practical land and sea observation applications, SAR images typically image large, complex scenes, and the targets to be detected are small, high-value, non-cooperative targets (such as unknown ships and vehicles). Observation times are limited, making it difficult to obtain sufficient training samples to support deep neural network learning. In such situations, feature extraction is required to analyze the characteristics of non-cooperative targets to achieve detection. However, strong scattering background objects can complicate detection, generating a large number of false alarms and reducing detection accuracy.

[0003] In the existing field of SAR image target detection, the most widely used traditional detection algorithm is the CFAR algorithm based on background clutter statistical modeling. The CFAR detection algorithm has the characteristics of constant false alarm rate and threshold adaptation, and has strong interpretability and robustness. However, with the development of SAR technology, image resolution has continued to increase, and background clutter characteristics have become more complex, which poses a challenge to traditional detection algorithms. Faced with multi-scale targets and non-uniform background clutter, the performance of traditional constant false alarm detection algorithms has declined. The traditional CFAR method detects targets pixel by pixel and cannot meet the requirements of real-time SAR image processing. Ordinary saliency algorithms indiscriminately treat all prominent parts in the image as salient objects, which will generate a large number of false alarms when performing saliency detection on SAR images of complex scenes containing a large number of man-made objects and natural background clutter.

[0004] Therefore, the present invention proposes a non-cooperative vehicle target detection method in SAR images based on collaborative saliency to solve the above technical problems. Summary of the Invention

[0005] To overcome the above-mentioned defects in the prior art, the present invention provides a method for non-cooperative vehicle target detection in SAR images based on collaborative saliency. The method uses auxiliary data and the data to be tested to calculate collaborative saliency to obtain the salient features of non-cooperative targets and complete detection. The collaborative saliency is applied to the SAR image target detection task. A preliminary collaborative saliency map is obtained by calculating the contrast features and joint features of the image. Finally, the weights generated from the collaborative saliency map of the auxiliary data are used to further suppress the background and enhance the target, thereby improving detection performance.

[0006] The above technical objectives of the present invention are achieved through the following technical solutions: a method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency, comprising the following steps:

[0007] S1, dividing a group image consisting of multiple single images into two parts: an auxiliary image and an image to be tested;

[0008] S2. Cluster all pixels on the group image using the K-Means clustering algorithm to obtain group image cluster pixel clusters. The specific steps of the K-Means clustering algorithm are as follows: set the number of pixel clusters to K in advance, randomly select K pixels as initial cluster centers, then calculate the distance between each pixel value and the cluster center of each pixel cluster, and assign each pixel to the cluster center closest to it;

[0009] S3, performing single image clustering on each single image to obtain single image clustered pixel clusters;

[0010] S4, extracting contrast features of the single image clustered pixel clusters, and obtaining a single image saliency map using contrast feature weights;

[0011] S5, extracting contrast features and joint features of the group image cluster pixel clusters, and calculating the image saliency map of the group using the contrast feature weights and the joint feature weights;

[0012] S6. Combining the single image saliency map and the group image saliency map to obtain a preliminary collaborative saliency map;

[0013] S7. Calculate the weight of the auxiliary image saliency map, and use the weight of the auxiliary image saliency map to weight the preliminary co-saliency map to obtain a final co-saliency map.

[0014] Furthermore, in step S2, each time a new pixel is assigned, the cluster center is recalculated based on the existing pixels in the cluster, and this process is repeated until the loss function corresponding to the clustering result is minimized. The loss function is calculated as follows:

[0015]

[0016] Among them, x i represents the i-th pixel, c i is x i The cluster to which it belongs, μ i represents the center point corresponding to the i-th cluster, and M represents the total number of pixels.

[0017] Furthermore, the calculation formula of the contrast feature weight in step S4 and step S5 is as follows:

[0018]

[0019] Among them, w c (k) represents the contrast feature weight; K represents the number of clusters; n i represents the number of all pixels in the i-th cluster; N represents the total number of all pixels; μ i Represents the center point corresponding to the i-th cluster.

[0020] Furthermore, the joint feature weight described in step S5 is expressed as cluster variance, and the calculation formula is as follows:

[0021]

[0022]

[0023] Among them, w d (k) represents the contrast feature weight; q k,j represents the proportion of the kth cluster in the jth image; n k,j N represents the total number of pixels in the kth cluster in the jth image; j Represents the total number of pixels in the jth image; M represents the total number of images.

[0024] Furthermore, in step S6, the weights of the single image saliency map and the group image saliency map are fused using a point-by-point multiplication method, and then the weight value of each pixel is calculated using a Gaussian distribution, and finally a preliminary collaborative saliency map is obtained;

[0025] The calculation formula of the point-by-point multiplication method is as follows:

[0026]

[0027] Among them, p(k) represents the weight value of the kth cluster; w i (k) represents the contrast feature weight and joint feature weight of the k-th cluster;

[0028] The calculation formula of the Gaussian distribution method is as follows:

[0029]

[0030]

[0031] Among them, p(x) represents the weight value of each pixel; v x represents the feature vector of pixel x; σ k represents the variance of the kth cluster.

[0032] Furthermore, the calculation formula for the weight of the auxiliary image saliency map in step S7 is as follows:

[0033] w r =max(p i (x)),i=1,…,N (8)

[0034] Among them, p i (x) represents the weight of the preliminary detection of co-saliency of the i-th auxiliary image; N represents the number of auxiliary images. r The final synergistic significance result is obtained by weighting the preliminary synergistic significance detection results.

[0035] In summary, the present invention has the following beneficial effects: the present invention applies collaborative saliency to the SAR image target detection task for the first time, improves the spatial position constraints of the original collaborative saliency algorithm according to the task characteristics, and finally uses the weights generated by the auxiliary data collaborative saliency map to further suppress the background and enhance the target, thereby improving the detection performance and providing a new solution to the non-cooperative target detection problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a flowchart of the method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency in Example 1 of the present invention;

[0037] Figure 2 This is a schematic diagram of joint feature description in Example 1 of the present invention;

[0038] Figure 3 This is a comparison diagram of the preliminary detection of collaborative saliency between the background slice and the target slice in Example 1 of the present invention;

[0039] Figure 4 This is the final effect diagram of the collaborative saliency of the background slice and the target slice in Example 1 of the present invention;

[0040] Figure 5 is the large scene image and its ground truth image in Example 2 of the present invention;

[0041] Figure 6 This is a trend diagram showing the effect of the amount of auxiliary data on the running time in Example 2 of the present invention;

[0042] Figure 7This is a trend diagram showing the effect of the amount of auxiliary data on the actual detection probability in Example 2 of the present invention;

[0043] Figure 8 This is a trend diagram showing the effect of the amount of auxiliary data on the actual false alarm probability in Example 2 of the present invention;

[0044] Figure 9 This is a graph showing the collaborative significance detection results in Example 2 of the present invention;

[0045] Figure 10 2 is a diagram of the detection results in Example 2 of the present invention;

[0046] Figure 11 1 is a diagram of the detection results after post-processing in Example 2 of the present invention;

[0047] Figure 12 This is the final result diagram in Example 2 of the present invention. DETAILED DESCRIPTION

[0048] The following is combined with Figure 1-12 The present invention is described in further detail.

[0049] Example 1:

[0050] like Figure 1 As shown in FIG, the non-cooperative vehicle target detection method in SAR images based on collaborative saliency includes the following steps:

[0051] S1, dividing a group image consisting of multiple single images into two parts: an auxiliary image and an image to be tested;

[0052] S2. Cluster all pixels on the group image using the K-Means clustering algorithm to obtain group image cluster pixel clusters. The specific steps of the K-Means clustering algorithm are as follows: set the number of pixel clusters to K in advance, randomly select K pixels as initial cluster centers, then calculate the distance between each pixel value and the cluster center of each pixel cluster, and assign each pixel to the cluster center closest to it;

[0053] S3, performing single image clustering on each single image to obtain single image clustered pixel clusters;

[0054] S4, extracting contrast features of the single image clustered pixel clusters, and obtaining a single image saliency map using contrast feature weights;

[0055] S5, extracting contrast features and joint features of the group image cluster pixel clusters, and calculating the image saliency map of the group using the contrast feature weights and the joint feature weights;

[0056] S6. Combining the single image saliency map and the group image saliency map to obtain a preliminary collaborative saliency map;

[0057] S7. Calculate the weight of the auxiliary image saliency map, and use the weight of the auxiliary image saliency map to weight the preliminary co-saliency map to obtain a final co-saliency map.

[0058] In step S2 above, clustering can organize pixels in an image that are similar in some way, thus uncovering the inherent structure between pixels. K-Means can divide data into K clusters, with the center of each cluster calculated by the mean of the values ​​within the cluster. The K-Means process involves setting the number of pixel clusters to K, randomly selecting K pixels as initial cluster centers, and then calculating the distance between each pixel value and the center of each cluster, assigning each pixel to the cluster center closest to it.

[0059] The cluster center and all pixels assigned to that cluster represent a cluster. Each time a new pixel is assigned, the cluster center is recalculated based on the existing pixels in the cluster. This process is repeated until the loss function corresponding to the clustering result is minimized. The loss function can be defined as the sum of squared errors between each pixel and the cluster center, which can be expressed as follows:

[0060]

[0061] Among them, x i represents the i-th pixel, c i is x i The cluster to which it belongs, μ i represents the center point corresponding to the i-th cluster, and M represents the total number of pixels.

[0062] The clustering mentioned in the present invention is divided into two parts. The first part is to cluster all images individually to obtain single image clustering results. The second part is to cluster all images as a whole to obtain group image clustering results.

[0063] The calculation formula for the contrast feature weight described in the above steps S4 and S5 is as follows:

[0064]

[0065] Among them, w c (k) represents the contrast feature weight; K represents the number of clusters; n i represents the number of all pixels in the i-th cluster; N represents the total number of all pixels; μ i Represents the center point corresponding to the i-th cluster.

[0066] In the above step S5, the joint feature is used to describe the distribution characteristics of pixels of the same category in multiple images. The frequency of pixels in the same cluster is an important description of the joint feature. Figure 2 As shown in the figure, the common pixels in the group image clustering are almost uniformly distributed. The top is the original image, the middle is the group clustering diagram, and the bottom is the histogram distribution of pixels from different clusters on a single image. The pixels in the group image are clustered into 10 clusters, as shown in the first row and second column of the histogram. They represent the clusters containing the vehicle target pixels, and it can be found that their distribution is roughly uniform. Although similar backgrounds also show a uniform distribution (as shown in the first column of the fourth row and the second column of the second row), these background areas are suppressed in the final co-significance map by fusing the contrast weights. Based on this, the cluster variance can be used to represent the joint feature weight, and the calculation formula is as follows:

[0067]

[0068]

[0069] Among them, w d (k) represents the contrast feature weight; q k,j represents the proportion of the kth cluster in the jth image; n k,j N represents the total number of pixels in the kth cluster in the jth image; j Represents the total number of pixels in the jth image; M represents the total number of images.

[0070] In the above step S6, common methods of fusion of significant features include linear summation or point-by-point multiplication. Multiplication has better noise suppression performance than summation. Summation has a higher recall rate than multiplication. However, for complex scene SAR images, a higher recall rate means an increase in false alarm rate. Therefore, the present invention adopts point-by-point multiplication to fuse the weights of the single image saliency map and the group image saliency map, and then uses Gaussian distribution to calculate the weight value of each pixel, and finally obtains a preliminary collaborative saliency map;

[0071] The calculation formula of the point-by-point multiplication method is as follows:

[0072]

[0073] Among them, p(k) represents the weight value of the kth cluster; w i (k) represents the contrast feature weight and joint feature weight of the k-th cluster;

[0074] The calculation formula of the Gaussian distribution method is as follows:

[0075]

[0076]

[0077] Among them, p(x) represents the weight value of each pixel; v x represents the feature vector of pixel x; σ k represents the variance of the kth cluster.

[0078] The calculation formula of the weight of the auxiliary image saliency map in step S7 is as follows:

[0079] w r =max(p i (x)),i=1,…,N (8)

[0080] Among them, p i (x) represents the weight of the preliminary detection of co-saliency of the i-th auxiliary image; N represents the number of auxiliary images. r The final synergistic significance result is obtained by weighting the preliminary synergistic significance detection results.

[0081] like Figure 3 The figure shows the preliminary results of the collaborative saliency test of background and target slices. It can be seen that if the target data is background, the saliency of the vehicle target in the auxiliary data is weakened; if the target data is a vehicle target, the saliency of the vehicle target in the auxiliary data is strengthened. Therefore, the saliency test results of the auxiliary data can be fed back to the target data, further suppressing the background and highlighting the target through a weighted approach.

[0082] like Figure 4 As shown in the figure, by using the saliency results of the auxiliary data to weight the feedback of the test data, it can be found that if the test data is the background, the saliency of the vehicle target in the auxiliary data is weak, and the background saliency is further suppressed after the feedback weighting; if the test data is the vehicle target, the saliency of the vehicle target in the auxiliary data is strong, and the vehicle target saliency is further enhanced after the feedback weighting.

[0083] Example 2:

[0084] The present invention uses the MSTAR dataset as auxiliary data and the X-band measured large scene data as the data to be tested. The size of each image in the MSTAR dataset is 128×128, and it contains 10 types of targets. Each type of target in the dataset has a large number of images with different azimuth angles, and the azimuth angle range is 0-180°, and the azimuth angle interval is about 1-2°. In addition, each type of data has two pitch angles of 17° and 15° to choose from. The large scene data contains several vehicle targets, none of which belong to the target category in the MSTAR dataset. The image size of the large scene data is 3000×3000, the resolution is 0.5 meters, the number of vehicle targets is 12, and the large scene image and its true value map are as follows: Figure 5In the experiment, the large scene data to be tested is cut into slices and collaborative saliency detection is performed with the MSTAR dataset to achieve the purpose of detecting vehicle targets in the large scene SAR image to be tested.

[0085] SAR uses the backscattering characteristics of objects to form images, which is very different from the imaging mechanism of optical images. Therefore, some targets with similar structures will have similar and strong backscattering coefficients, and some ground objects similar to natural vegetation have medium to weak backscattering coefficients. After SAR imaging, the target brightness is generally strong and there is a strong contrast between the target and the background. The intensity of the auxiliary data target is similar to the intensity of the target in the test data. Using collaborative saliency to detect the target of the test data can effectively improve the detection performance, but increasing the number of auxiliary data will reduce the efficiency of collaborative saliency detection. Therefore, it is crucial to reasonably select the number of auxiliary data to improve the efficiency of collaborative saliency target detection. In this embodiment, N (N = 1, ..., 10) auxiliary data are randomly selected from the MSTAR data set to analyze the impact of the number of auxiliary data on the performance of collaborative saliency detection.

[0086] In order to quantitatively evaluate the detection results, the generated co-saliency map is binarized and compared with the ground-truth map, and the following formula is used for quantitative evaluation:

[0087]

[0088]

[0089] Among them, P d represents the actual detection probability; P f Indicates the actual false alarm probability; N t Indicates the number of target pixels of the vehicle to be tested in the image; N c Indicates the number of clutter pixels in the image; N dt Indicates the number of pixels of correctly detected vehicle targets; N dc Indicates the number of pixels of background clutter detected.

[0090] Figure 6 This is a trend chart showing the impact of the amount of auxiliary data on the running time. It can be seen that as the amount of auxiliary data increases, the running time basically increases linearly.

[0091] Figure 7 Figure 2 is a trend chart showing the impact of the number of auxiliary data on the actual detection probability. It can be seen that when the number of auxiliary data N is 1 to 4, the actual detection probability of collaborative saliency detection shows a slowly increasing trend. However, when the number of auxiliary data N is 5 to 10, the actual detection probability of collaborative saliency detection shows a basically flat trend.

[0092] Figure 8This is a trend chart showing the impact of the amount of auxiliary data on the actual false alarm probability. It can be seen that as the amount of auxiliary data increases, the actual false alarm probability slowly increases. The auxiliary data also contains some background, including some coherent speckle noise. As the amount of auxiliary data increases, the probability of coherent speckle noise appearing in co-saliency detection increases, and it will also be detected as part of the co-saliency.

[0093] Figure 9 The co-saliency detection result graphs with different amounts of auxiliary data are given.

[0094] The running time, actual detection probability, and actual false alarm probability under different amounts of auxiliary data are shown in Table 1 below:

[0095] Table 1 Analysis of the impact of the amount of auxiliary data on the performance of collaborative saliency detection

[0096] N Runtime <![CDATA[P d ]]> <![CDATA[P f ]]> 1 38.28s 0.6239 0.0035 2 58.98s 0.6121 0.0032 3 85.58s 0.6447 0.0046 4 102.25s 0.6779 0.0067 5 119.75s 0.6496 0.0072 6 163.33s 0.6493 0.0087 7 232.73s 0.6441 0.0087 8 307.14s 0.6475 0.0094 9 416.13s 0.6495 0.0095 10 495.81s 0.6572 0.0099

[0097] In practical applications, the optimal amount of auxiliary data depends on the specific problem and scenario being solved. In this embodiment, the algorithm only calculates simple contrast and joint features, so only a small amount of data is needed to extract the basic patterns in the data. Increasing the amount of auxiliary data does not significantly improve the actual detection probability. Increasing the amount of auxiliary data will lead to a decrease in detection efficiency and a significant increase in the actual false alarm probability. Therefore, to address the problem of reducing false alarms while maintaining detection rate, the amount of auxiliary data in this embodiment is set to N = 2.

[0098] In the MSTAR dataset, objects of the same category have similar morphology. In this example, two images are randomly selected from the MSTAR dataset as auxiliary data. Figure 5 Co-saliency detection is performed on the large scene image shown in (a).

[0099] In this embodiment, the detection results of the algorithm in this paper are compared with those of the traditional CA-CFAR detector, SO-CFAR detector, GO-CFAR detector, RS saliency algorithm, ITTI saliency algorithm and FineGrained saliency algorithm to prove the effectiveness of the algorithm in this embodiment. The constant false alarm rate P of CA-CFAR detector, SO-CFAR detector and GO-CFAR detector is set to fa =0.05, the test results are as follows Figure 10 shown.

[0100] Depend on Figure 10As can be seen, the CA-CFAR algorithm misses many targets, with many gaps between target parts; the GO-CFAR algorithm produces a large number of missed detections, with the target area largely undetected; and the SO-CFAR algorithm, the RS saliency algorithm, the ITTI saliency algorithm, and the FineGrained saliency algorithm detect most targets but generate a large number of false alarms. The collaborative saliency detection algorithm proposed in this paper can effectively suppress false alarms while maintaining detection rates, thus extracting vehicle targets.

[0101] Table 2 below provides a quantitative comparison of detection performance. As can be seen from Table 2, the algorithm proposed in this paper has a significant advantage in detection speed compared to the traditional CFAR algorithm. The traditional CFAR algorithm requires pixel-by-pixel statistics, which takes a lot of time to calculate in large scene images. The algorithm proposed in this paper only needs to divide the image to be tested into slices and calculate the results together with auxiliary data, which can greatly improve the detection efficiency of the target. Compared with the CFAR algorithm, the calculation time of this algorithm is improved by at least 1500.85 seconds. The algorithm proposed in this paper has the lowest false alarm rate and has a significant advantage in suppressing false alarms compared to other algorithms. Compared with other algorithms, the actual false alarm probability of the algorithm proposed in this paper has decreased by at least 38.5%.

[0102] Table 2 Algorithm performance comparison

[0103] algorithm Runtime <![CDATA[P d ]]> <![CDATA[P f ]]> Co-significance algorithm of the present invention 58.98s 0.6121 0.0032 CA-CFAR 1599.83s 0.2235 0.0127 GO-CFAR 1945.37s 0.0001 0.0052 SO-CFAR 1957.92s 0.7838 0.0608 RS saliency algorithm 17.48s 0.9717 0.1633 ITTI saliency algorithm 2.92s 0.9915 0.1312 FineGrained Saliency Algorithm 1.52s 0.8196 0.1087

[0104] right Figure 10 The results shown in the figure are first closed to connect the target broken areas, then area filtering is performed to leave areas with an area size similar to the vehicle target, and finally length and width filtering is performed to suppress the road edges and clutter that exceed the length and width limits, and the detection result diagram after post-processing is obtained, as shown in the figure below. Figure 11 shown.

[0105] use Figure 11 The performance comparison table after drawing post-processing is shown in Table 3 below:

[0106] Table 3 Performance comparison after post-processing

[0107] algorithm <![CDATA[P d ]]> <![CDATA[P f ]]> Whether the target is missed Co-significance algorithm of the present invention 0.6098 0.0001 no CA-CFAR 0.0648 0.0008 yes GO-CFAR 0.0001 0.0008 yes SO-CFAR 0.3576 0.0059 yes RS saliency algorithm 0.8305 0.0036 yes ITTI saliency algorithm 0.9808 0.0050 no FineGrained Saliency Algorithm 0.2871 0.0006 yes

[0108] As shown in Table 3, except for the proposed algorithm and the ITTI algorithm, all other algorithms experience missed target detection. The ITTI algorithm also has a high false alarm rate and cannot effectively distinguish between targets and clutter. The proposed algorithm can reduce false alarms while ensuring target detection. Compared to other algorithms, the proposed algorithm reduces the actual false alarm rate by at least 83.3%.

[0109] In this embodiment, in order to intuitively compare the detection performance of the model, Figure 11 The connected domain of the detection result shown is framed by the target detection frame, and the result is as follows Figure 12 As shown in the figure, F1-score is used as the performance evaluation indicator of the detection results. F1-score is a harmonic average of the model's precision and recall, with a value range of [0,1]. The calculation formula for F1-score is:

[0110]

[0111]

[0112]

[0113] Among them, TP represents the number of correct targets in the detection results, FP represents the number of false targets, NP represents the number of true targets, P represents the detection precision, R represents the recall rate, and F1-score represents the main indicator for evaluating performance.

[0114] The final performance comparison table, shown in Table 4 below, shows that our algorithm outperforms the CA-CFAR, SO-CFAR, GO-CFAR, RS, ITTI, and FineGrained algorithms in detection performance. The CA-CFAR, SO-CFAR, GO-CFAR, RS, and FineGrained algorithms all miss detections and have weak false alarm removal capabilities. While the TTI algorithm does not miss any targets, its false alarm removal capability is weaker than our algorithm. Compared to other algorithms, our algorithm improves detection accuracy by at least 0.52 and F1-score by at least 0.49.

[0115] Table 4 Final performance comparison table

[0116] algorithm TP FP P R F1-score Co-significance algorithm of the present invention 12 4 0.75 1.00 0.86 CA-CFAR 4 446 0.01 0.33 0.02 GO-CFAR 0 363 0.00 0.00 0 SO-CFAR 6 289 0.02 0.50 0.04 RS saliency algorithm 10 34 0.23 0.83 0.36 ITTI saliency algorithm 12 41 0.23 1.00 0.37 FineGrained Saliency Algorithm 4 162 0.02 0.33 0.04

[0117] In summary, the present invention proposes a non-cooperative vehicle target detection method based on collaborative saliency, which regards existing slice data as auxiliary data to assist the data to be tested in vehicle target detection. By extracting targets in the data to be tested that have similar salience to the auxiliary data to complete the detection task, the algorithm can reduce the false alarm rate while ensuring the detection rate. In the experiment, MSTAR data was used as auxiliary data and 3000×3000 large scene measured data was used as the data to be tested to verify the effectiveness of the algorithm. The present invention applies collaborative saliency to the SAR image target detection task for the first time, improves the spatial position constraints of the original collaborative saliency algorithm, and finally uses the weights generated by the auxiliary data collaborative saliency map to further suppress the background and enhance the target, thereby improving the detection performance.

[0118] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A non-cooperative vehicle target detection method in SAR images based on collaborative saliency, characterized by: The following steps are involved: S1, dividing a group image consisting of multiple single images into two parts: an auxiliary image and an image to be tested; S2. Cluster all pixels on the group image using the K-Means clustering algorithm to obtain group image cluster pixel clusters. The specific steps of the K-Means clustering algorithm are as follows: set the number of pixel clusters to K in advance, randomly select K pixels as initial cluster centers, then calculate the distance between each pixel value and the cluster center of each pixel cluster, and assign each pixel to the cluster center closest to it; S3, performing single image clustering on each single image to obtain single image clustered pixel clusters; S4, extracting contrast features of the single image clustered pixel clusters, and obtaining a single image saliency map using contrast feature weights; S5, extracting contrast features and joint features of the group image cluster pixel clusters, and calculating the image saliency map of the group using the contrast feature weights and the joint feature weights; S6. Combining the single image saliency map and the group image saliency map to obtain a preliminary collaborative saliency map; S7. Calculate the weight of the auxiliary image saliency map, and use the weight of the auxiliary image saliency map to weight the preliminary co-saliency map to obtain a final co-saliency map.

2. The method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency according to claim 1, characterized in that: In step S2, each time a new pixel is assigned, the cluster center is recalculated based on the existing pixels in the cluster, and this process is repeated until the loss function corresponding to the clustering result is minimized. The loss function is calculated as follows: Among them, x i represents the i-th pixel, c i is x i The cluster to which it belongs, μ i represents the center point corresponding to the i-th cluster, and M represents the total number of pixels.

3. The method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency according to claim 1, wherein: The calculation formula of the contrast feature weight described in step S4 and step S5 is as follows: Among them, w c (k) represents the contrast feature weight; K represents the number of clusters; n i represents the number of all pixels in the i-th cluster; N represents the total number of all pixels; μ i Represents the center point corresponding to the i-th cluster.

4. The method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency according to claim 1, wherein: The joint feature weight described in step S5 is expressed as cluster variance, and the calculation formula is as follows: Among them, w d (k) represents the contrast feature weight; q k,j represents the proportion of the kth cluster in the jth image; n k,j N represents the total number of pixels in the kth cluster in the jth image; j Represents the total number of pixels in the jth image; M represents the total number of images.

5. The method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency according to claim 1, wherein: In step S6, the weights of the single image saliency map and the group image saliency map are fused using a point-by-point multiplication method, and then the weight value of each pixel is calculated using a Gaussian distribution, and finally a preliminary collaborative saliency map is obtained; The calculation formula of the point-by-point multiplication method is as follows: Among them, p(k) represents the weight value of the kth cluster; w i (k) represents the contrast feature weight and joint feature weight of the k-th cluster; The calculation formula of the Gaussian distribution method is as follows: Among them, p(x) represents the weight value of each pixel; v x represents the feature vector of pixel x; σ k represents the variance of the kth cluster.

6. The method for detecting non-cooperative vehicle targets in SAR images based on collaborative saliency according to claim 1, wherein: The calculation formula of the weight of the auxiliary image saliency map in step S7 is as follows: w r =max(p i (x)),i=1,…,N (8) Among them, p i (x) represents the weight of the preliminary detection of co-saliency of the i-th auxiliary image; N represents the number of auxiliary images, using w r The final synergistic significance result is obtained by weighting the preliminary synergistic significance detection results.

Images