CN113378763B - SAR image-oriented target automatic detection method and system - Google Patents

Abstract

The invention discloses an automatic target detection method and system for SAR images, including inputting an input image into a pre-trained feature fusion and anti-interference network FFAN to obtain a target frame for target detection, and the feature fusion and anti-interference network FFAN includes : Backbone network, pooling and channel shuffling module, deep fusion module, spatial attention mechanism RSAM, spatial feature fusion module ASFF and classification regression network. The invention can effectively extract and fuse features and improve regression accuracy, so that the feature fusion and anti-interference network FFAN can significantly improve the detection effect under the premise that the test speed is hardly affected. The method of the invention has the advantage of good versatility, not only can be used to realize high-precision and fast detection of aircraft, but also can be used for detection and recognition of other dense targets in SAR images.

Description

A method and system for automatic target detection in SAR images

technical field

The invention relates to a target detection technology based on SAR images, in particular to an automatic target detection method and system for SAR images.

Background technique

Target detection is a hot issue in the field of high-resolution synthetic aperture radar (SAR). Aircraft is a typical man-made target. In the military field, aircraft has important military value. Efficient and accurate acquisition of aircraft target information in airports, airspace and other locations is of great significance for real-time acquisition of battlefield military intelligence and formulation of combat plans; In the field, the detection and identification of aircraft can manage the airport more effectively. Because Synthetic Aperture Radar (SAR) has the characteristics of all-day and all-weather work, the target detection and recognition technology in SAR images has been widely concerned and researched.

Traditional SAR image target detection methods, such as Constant False Alarm Rate (CFAR) and CA-CFAR algorithms, have problems such as poor robustness and low degree of automation that are difficult to break through.

In recent years, synthetic aperture radar imaging and deep learning technology have developed rapidly, and people have obtained more high-resolution SAR images. Many scholars have begun to study the use of deep learning algorithms for SAR image target detection. Compared with traditional SAR image target detection methods and machine learning algorithms, it has higher detection accuracy and shorter test time, and has made great research progress, especially end-to-end detection can be realized. Aircraft targets have the characteristics of small size, various types, and dense distribution, which is more difficult than target detection such as runways, bridges and buildings. Therefore, target detection is a difficult point in SAR image target detection. In 2019, Li et al. combined the improved line segment detector LSD with Faster-CNN [4] to design an aircraft detection method in SAR images. In order to better achieve target detection, more and more scholars have introduced the attention mechanism. Zhao et al. designed a pyramid attention dilated network pyramid attention dilated network (PADN) to enhance the learning of aircraft's backscatter fine-grained features. In 2020, Guo et al. proposed a method for detecting aircraft by combining scattering information with deep learning to detect aircraft targets. Chen et al. proposed a multi-level densely connected dual-attention network to automatically detect airport runway areas, and achieved better extraction results, but the network training speed and testing speed were relatively slow. In 2020, Chen et al. proposed a new end-to-end framework based on deep learning to automatically classify water and shadow areas in SAR images. Chen et al. proposed a new framework for scene classification called Feature Recalibration Network with Multi-Scale Spatial Features (FRN-MSF), which achieved high-precision classification for SAR scenes. Tan et al. proposed a geospatial contextual attention mechanism GCAM to automatically detect the airport area, which not only has high accuracy, but also has a short test time. In 2021, Wang et al. combined the GCAM algorithm with the aircraft detection algorithm and proposed an efficient SAR image aircraft detection framework, which realized automatic and rapid detection of aircraft targets.

EWFAN is an efficient SAR image target detection network, which has considerable advantages in both accuracy and speed. However, due to the complex SAR image background and various types of aircraft, there are still problems in using the EWFAN algorithm to detect aircraft, and the detection rate and false alarm rate have not reached a satisfactory level.

Contents of the invention

The technical problem to be solved by the present invention: aiming at the above-mentioned problems of the prior art, a kind of

In order to solve the problems of the technologies described above, the technical solution adopted in the present invention is:

A target automatic detection method for SAR images, comprising inputting an input image into a pre-trained feature fusion and anti-jamming network FFAN to obtain a target frame for target detection, the feature fusion and anti-jamming network FFAN comprising:

The backbone network is used to downsample the input image into features of different scales, including low-level features P3, intermediate-level features P4-P6 and high-level features P7;

The pooling and channel shuffling modules are used to extract and fuse the intermediate-level features P4~P6 respectively to obtain the corresponding shuffling features P4'~P6';

The deep fusion module is used to perform deep fusion on the low-level features P3, shuffled features P4′～P6′, and high-level features P7, and respectively obtain the corresponding deep-fused low-level features P3′, shuffled features P4″～P6″, high-level features Feature P7';

The spatial attention mechanism RSAM is used to further extract and fuse the shuffled features P6″ and advanced features P7′ after deep fusion;

The spatial feature fusion module ASFF is used for further spatial feature extraction and fusion of the deep-fused low-level features P3′ and shuffled features P4″～P5″ respectively;

The classification and regression network is used to extract and fuse all the spatial features output by the spatial attention mechanism RSAM and the spatial feature fusion module ASFF for classification and regression to generate a target frame for target detection.

Optionally, the backbone network is an Efficientnet network, and the Efficientnet network sequentially performs down-sampling of different sizes on the input image of 512×512 to obtain corresponding features P1-P7 respectively, and the 3rd to 7th times to the next The features obtained by sampling are respectively regarded as low-level features P3, intermediate-level features P4-P6, and high-level features P7.

Optionally, the pooling and channel shuffling modules include a total of three shuffling modules, the shuffling module PSM-1, the shuffling module PSM-m, and the shuffling module PSM-s, and the shuffling module PSM-1 is used for The intermediate-level feature P4 is extracted and fused to obtain the corresponding shuffling feature P4′, and the shuffling module PSM-m is used to extract and fuse the intermediate-level feature P5 to obtain the corresponding shuffling feature P5′, and the shuffling module PSM- s is used to extract and fuse the intermediate-level features P6 to obtain corresponding shuffling features P6' respectively, and the shuffling module PSM-1, the shuffling module PSM-m, and the shuffling module PSM-s all include:

The channel separation module is used to divide the channel of the input feature into two;

The maximum pooling operation module is used to perform a maximum pooling operation on a part of the channels output by the channel separation module;

The average pooling operation module is used to perform an average pooling operation on another part of the channels output by the channel separation module;

A splicing module for splicing the output of the maximum pooling operation module, the output of the average pooling operation module and the original input feature of the input channel separation module;

The shuffling module is used for shuffling the output of the splicing module to enrich the information of each channel;

The convolution module is used to perform a convolution operation on the output of the shuffling module so that the number of output feature channels is the same as the input feature;

The pooling kernel sizes of the shuffling module PSM-1, the shuffling module PSM-m, and the shuffling module PSM-s for the maximum pooling operation and the average pooling operation are respectively {9, 15, 25}, { 5,11} and {3}.

Optionally, the deep fusion module is an improved structure of BiFPN, which adds four skip connections on the basis of BiFPN, the first jump connects the intermediate-level feature P6 and the shuffling feature P5′, and the second jump connects The intermediate-level feature P5 and the shuffling feature P4', the third jump connects the intermediate-level feature P5 and the intermediate-level feature P7, and the fourth jump connects the intermediate-level feature P4 and the intermediate-level feature P6.

Optionally, the classification regression network includes a classification network and a regression network, the classification network uses three convolutions of 64 channels and one convolution of n_b×n_c to predict the category of each target frame, and the regression network Three 64-channel convolutions and one n_b×4 convolution are used to predict the regression of each prior frame, where n_b refers to the number of prior frames owned by the feature layer, and n_c refers to Number of categories for network object detection.

Optionally, before inputting the input image into the pre-trained feature fusion and anti-jamming network FFAN to obtain the target frame of the target detection, it also includes the step of pre-training the feature fusion and anti-jamming network FFAN, and training the feature fusion and anti-jamming network The loss function _LAIF used in FFAN is:

L _AIF =L _FL +L _AIoU ,

In the above formula, L _FL is Focal loss classification loss, L _AIoU is AIoU regression loss, and there are:

L _AIoU = L _CIoU + pw,

In the above formula, L _CIoU is the CIoU regression loss, a and a ^t represent the area of the predicted frame and the real frame, respectively, p is the weight function of the area, w is the similarity of the area, and IoU is the intersection and union ratio.

In addition, the present invention also provides an automatic target detection method for SAR images, including:

S1) The input SAR image is generated as a window image in the form of a sliding window with overlapping regions;

S2) using each window image as an input image, using the steps of the SAR image-oriented target automatic detection method to perform target detection to obtain a target frame;

S3) converting the target frame coordinates of each window image into the original SAR image;

S4) Filter the overlapping target frames in the original SAR image to obtain the target frame as the final target detection result.

Optionally, the method used to filter the overlapping target frames in the original SAR image in step S4) is a non-maximum value suppression method.

In addition, the present invention also provides a SAR image-oriented target automatic detection system, which includes interconnected microprocessors and memory, the microprocessor is programmed or configured to execute the steps of the SAR image-oriented target automatic detection method.

In addition, the present invention also provides a computer-readable storage medium, where a computer program programmed or configured to execute the SAR image-oriented automatic target detection method is stored.

Compared with the prior art, the present invention has the following advantages:

1. The feature fusion and anti-interference network FFAN adopted by the method of the present invention includes: a backbone network for down-sampling the input image into features of different scales, including low-level features P3, intermediate-level features P4-P6, and high-level features P7; pooling The channel shuffling module is used to extract and fuse the intermediate-level features P4~P6 to obtain the corresponding shuffled features P4'~P6' respectively; the deep fusion module is used to extract and fuse the low-level features P3 and shuffled features P4'~ P6' and advanced features P7 are deeply fused to obtain the corresponding low-level features P3', shuffled features P4"~P6", and high-level features P7' after deep fusion respectively; the spatial attention mechanism RSAM is used to combine deep fusion The shuffled features P6″ and high-level features P7′ are used for further spatial feature extraction and fusion; the spatial feature fusion module ASFF is used to further further integrate the low-level features P3′ and shuffled features P4″~P5″ after deep fusion. Spatial feature extraction and fusion; classification and regression network, used to extract and fuse all spatial features output by the spatial attention mechanism RSAM and spatial feature fusion module ASFF for classification and regression to generate target frames for target detection.Through the above structure, it can be effectively The extraction and fusion of features and the improvement of regression accuracy make the feature fusion and anti-interference network FFAN significantly improve the detection effect without affecting the test speed.

2. The method of the present invention has the advantage of good versatility, and can not only be used to realize high-precision and fast detection of aircraft, but also can be used for detection and recognition of other dense targets in SAR images.

Description of drawings

FIG. 1 is a schematic structural diagram of a feature fusion and anti-interference network FFAN in an embodiment of the present invention.

FIG. 2 is a schematic diagram of an application arrangement method of feature fusion and anti-interference network FFAN in an embodiment of the present invention.

Fig. 3 is a schematic structural diagram of a pooling and channel shuffling module in an embodiment of the present invention.

Fig. 4 is a schematic diagram of structural improvement of the deep fusion module in the embodiment of the present invention.

Fig. 5 is a schematic structural diagram of the classification and regression network in the embodiment of the present invention

Fig. 6 is a principle analysis diagram of problems existing in CIoU Loss in the embodiment of the present invention.

Detailed ways

As shown in Figure 1, the automatic target detection method for SAR images in this embodiment includes inputting the input image into the pre-trained feature fusion and anti-interference network FFAN (Feature Fusion and Anti-interference Network) to obtain the target frame for target detection, and the feature fusion FFAN with anti-jamming network includes:

The backbone network is used to downsample the input image into features of different scales, including low-level features P3, intermediate-level features P4-P6 and high-level features P7;

The pooling and channel shuffling modules are used to extract and fuse the intermediate-level features P4~P6 respectively to obtain the corresponding shuffling features P4'~P6';

The deep fusion module is used to perform deep fusion on the low-level features P3, shuffled features P4′～P6′, and high-level features P7, and respectively obtain the corresponding deep-fused low-level features P3′, shuffled features P4″～P6″, high-level features Feature P7';

The spatial attention mechanism RSAM is used to further extract and fuse the shuffled features P6″ and advanced features P7′ after deep fusion;

The spatial feature fusion module ASFF is used for further spatial feature extraction and fusion of the deep-fused low-level features P3′ and shuffled features P4″～P5″ respectively;

The classification and regression network is used to extract and fuse all the spatial features output by the spatial attention mechanism RSAM and the spatial feature fusion module ASFF for classification and regression to generate a target frame for target detection.

Referring to Fig. 1, the backbone network in this embodiment is the Efficientnet network, and the Efficientnet network performs successive downsampling on the input image of 512×512 to obtain the corresponding features P1-P7 respectively, and the 3rd to 7th times to the next The features obtained by sampling are respectively regarded as low-level features P3, intermediate-level features P4-P6, and high-level features P7. The Efficientnet network was proposed by the Google team. It uses MBConv in MobileNet V2 as the main component of the model, and introduces the squeeze and excitation (SE) method in SENet to adjust the network structure. The Efficientnet network has certain advantages in terms of accuracy and speed, and is suitable as a backbone network for efficient target detection algorithms.

Pooling and channel shuffling modules are used to further extract effective features. As shown in Figure 3, the pooling and channel shuffling modules in this embodiment include three shuffling modules: shuffling module PSM-1, shuffling module PSM-m, shuffling module PSM-s, and shuffling module PSM- l is used to extract and fuse the intermediate-level features P4 to obtain the corresponding shuffled features P4′, and the shuffled module PSM-m is used to extract and fuse the intermediate-level features P5 to obtain the corresponding shuffled features P5′ respectively. The shuffling module PSM-s is used to extract and fuse the intermediate-level features P6 to obtain the corresponding shuffling features P6′ respectively, as shown in Figure 3, the shuffling module PSM-1, the shuffling module PSM-m, and the shuffling module PSM -s both include:

The channel separation module is used to divide the channel of the input feature into two;

The maximum pooling operation module is used to perform a maximum pooling operation on a part of the channels output by the channel separation module;

The average pooling operation module is used to perform an average pooling operation on another part of the channels output by the channel separation module;

A splicing module for splicing the output of the maximum pooling operation module, the output of the average pooling operation module and the original input feature of the input channel separation module;

The shuffling module is used for shuffling the output of the splicing module to enrich the information of each channel;

The convolution module is used to perform a convolution operation on the output of the shuffling module so that the number of output feature channels is the same as the input feature;

The pooling kernel sizes of the shuffling module PSM-l, the shuffling module PSM-m, and the shuffling module PSM-s for the maximum pooling operation and the average pooling operation are {9,15,25}, {5, 11} and {3}.

The deep fusion module is used to deeply fuse the features of the five-layer feature map to promote the information exchange between the feature layer and the channel to improve the detection effect. As shown in Figure 4, the deep fusion module is an improved structure of BiFPN, which adds four jump connections on the basis of BiFPN (as shown by the dotted arrow in the sub-figure (c) in Figure 4), the first jump Connect the intermediate level feature P6 and the shuffled feature P5′, the second jump connects the intermediate level feature P5 and the shuffled feature P4′, the third jump connects the intermediate level feature P5 and the intermediate level feature P7, and the fourth jump connects the intermediate level Feature P4 and intermediate level feature P6. Referring to FIG. 4 , the deep fusion module in this embodiment is recorded as M-BiFPN, M-BiFPN is an improvement of BiFPN, and BiFPN is an improvement of FPN. BiFPN (as shown in Figure 4(b)) not only has top-down fusion, but also bottom-up fusion, and BiFPN removes points with only one input source, compared with FPN (as shown in Figure 4(a)) , the accuracy and efficiency have been improved to a certain extent. In this embodiment, four skip connections are added on the basis of BiFPN, and the original feature map of each layer is added to the output, so that more features can be fused to improve network performance.

On the basis of the preliminary feature fusion completed by the deep fusion module, in this embodiment, the spatial attention mechanism RSAM and the spatial feature fusion module ASFF are used for further spatial feature extraction and fusion. The spatial attention mechanism RSAM performs further spatial feature extraction and fusion on the deeply fused shuffling features P6″ and high-level features P7′; the spatial feature fusion module ASFF takes the deeply fused low-level features P3′ and shuffling features P4″～ P5″ for further spatial feature extraction and fusion. Using the spatial attention mechanism RSAM and the spatial feature fusion module ASFF for further spatial feature extraction and fusion, they can enhance the saliency of the target and effectively suppress the interference of negative samples, and test the speed of the network little impact.

In the end, the feature map is further extracted and fused by using the spatial attention mechanism RSAM and the spatial feature fusion module ASFF to obtain five effective feature layers. The feature fusion and anti-interference network FFAN uses the classification and regression network to predict the results. As shown in Figure 5, the classification regression network includes a classification network and a regression network. The classification network uses three 64-channel convolutions and one n_b×n_c convolution to predict the category of each target frame, and the regression network uses three 64-channel convolutions. Channel convolution and 1 n_b×4 convolution are used to predict the regression of each prior frame, where n_b refers to the number of prior frames owned by the feature layer, and n_c refers to the network target detection category number.

Before inputting the input image into the pre-trained feature fusion and anti-interference network FFAN to obtain the target frame of the target detection, the step of pre-training feature fusion and anti-interference network FFAN is also included. In this embodiment, CIoU Loss is improved, and then AIoU Loss is introduced into Focal Loss, and AIF Loss is proposed. The loss function _LAIF used in training feature fusion and anti-interference network FFAN is:

L _AIF =L _FL +L _AIoU ,

In the above formula, L _FL is Focal loss classification loss, L _AIoU is AIoU regression loss, and there are:

L _AIoU = L _CIoU + pw,

In the above formula, L _CIoU is the CIoU regression loss, a and a ^t represent the area of the predicted frame and the real frame, respectively, p is the weight function of the area, w is the similarity of the area, and IoU is the intersection and union ratio. CIoU Loss adds an aspect ratio penalty item to DIoU Loss. This improvement can improve the network regression accuracy to a certain extent. However, when there is a common center point and the aspect ratio of the predicted box and the real box are the same, CIoU Loss will also degenerate into IoU Loss, as shown in Figure 6. Therefore, the loss function _LAIF used in training feature fusion and anti-jamming network FFAN in this embodiment designs a penalty item pw, which takes into account the difference between the area of the predicted frame and the real frame, which can make the area of the predicted frame closer to The area of the ground truth box. As shown in Figure 6, compared with the prediction frame 1, the area of the prediction frame 2 is closer to the area of the prediction frame 1 real frame, the L _CIoU of the prediction frame 2 is equal to the L _CIoU of the prediction frame 1, and the L _AIoU of the prediction frame 1 is larger than the prediction frame 2 L _AIoU .

Referring to Fig. 2, the present embodiment also provides a method for automatic target detection of SAR images, including:

S1) The input SAR image is generated as a window image in the form of a sliding window with overlapping regions;

S2) Using each window image as an input image, using the steps of the aforementioned SAR image-oriented target automatic detection method of this embodiment to perform target detection to obtain a target frame;

S3) converting the target frame coordinates of each window image into the original SAR image;

S4) Filter the overlapping target frames in the original SAR image to obtain the target frame as the final target detection result.

In step S1) of this embodiment, when the window image is generated in the form of a sliding window with overlapping regions, adjacent windows will overlap by 20%; since the sliding window detection has a certain overlapping rate, overlapping frames will be generated, so Step S4) It is necessary to filter the overlapping target frames in the original SAR image to obtain the target frame as the final target detection result. Referring to Fig. 2, the method used to filter the overlapping target frames in the original SAR image in step S4) of this embodiment is the non-maximum value suppression method (NMS algorithm). Undoubtedly, other known filtering methods can also be used as needed or algorithm. In the detection process, the feature fusion and anti-interference network FFAN will generate multiple predicted target frames with different confidence levels on one target. In this embodiment, the NMS algorithm is used to remove these redundant candidate frames and retain the frame that best matches the target. This can effectively reduce the number of false alarms.

In the following, a verification experiment will be conducted on the feature fusion and anti-jamming network FFAN in the method of this embodiment. In the experiment, more than a dozen airport images of 1m resolution of the Gaofen-3 system were used. The data set contains 4396 aircraft samples with a size of 500×500 and the corresponding label files. The ratio of the training set to the verification set is 4:1. In addition, in this embodiment, an independent test experiment of aircraft detection was carried out using SAR images not used in the dataset. The experiments in this example are all carried out under the pytorch framework of the Ubuntu 18.04 system. All networks are trained using the same dataset. Each model is trained on two RTX2080Ti GPUs for 100 epochs, with an initial learning rate of 0.0005 and a batch size of 16. The input image size is 512×512 pixels. For the sake of fairness, none of the four models Efficientdet, YOLOV4, EWFAN, and FFAN in this example use pre-training models and automatic data enhancement functions, and do not use any advanced testing techniques such as Softer-NMS or data enhancement etc. During the testing phase, the NMS thresholds of the four models are all set to 0.5.

In this embodiment, two evaluation indicators, detection rate (DR) and false alarm rate (FAR), are used to measure the effect of the network, and the specific calculation formula is as follows:

Among them, the detection rate represents the ratio of the number of aircraft targets correctly detected by the network (C) to the number of aircraft targets in the label (L), and the false alarm rate is the ratio of the number of false alarms to the number of prediction frames output by the network (S).

In order to verify the performance of the proposed algorithm, in this example, the airport I (Hongqiao Airport, 12000×14400) and the airport II (Capital Airport, 14400×16800) in the Gaofen-3 system were independently tested, and the aircraft detection performance Analytical evaluation was carried out. The test results of the aircraft are shown in Table 1.

Table 1: Comparison of aircraft detection results in SAR images of different networks.

Both Airport I and Airport II are large civil airports. The aircraft targets in the airport have the characteristics of many types, small size, dense distribution, etc., and the image background is complex, which increases the difficulty of detection. Airport I and Airport II have 127 and 143 aircraft respectively. Based on the analysis of the detection results of the two airports, EfficientDet has the worst detection effect. Its false alarm rate is very high, with an average false alarm rate of 40.58%, and the average detection rate is also low. Compared with EfficientDet, the false alarm rates of YOLOv4 and EWFAN are slightly lower, 31.95% and 18.20%, respectively, but the false alarm rates are still high. FFAN (this embodiment) has the best detection effect, its average detection rate and false alarm rate are 86.26% and 8.42%, respectively, which shows that the network can extract features well and improve the ability to resist background interference. From the perspective of test time, the average test time of FFAN (this embodiment) is 23.87s, which is only 0.83s higher than that of EfficientDet. Therefore, FFAN (this embodiment) greatly improves the overall test speed while the test speed is basically unaffected. The detection effect has great engineering application value.

To sum up, target detection is a hot issue in the field of high-resolution synthetic aperture radar (SAR). Aiming at the problems of complex background and diverse features of aircraft targets in SAR images, this embodiment proposes an efficient automatic detection method for aircraft targets , this method is based on feature fusion and anti-interference network FFAN, which can more effectively extract and fuse aircraft features, and quickly obtain satisfactory detection results. Its detection rate and false alarm rate are superior to existing EfficientDet, EWFAN, YOLOv4 and other aircraft detection methods . This method based on feature fusion and anti-jamming network FFAN can be used for SAR image target detection, and can achieve satisfactory detection results in a short test time, and can also be used for the detection and recognition of other dense targets in SAR images. It has practical application value.

In addition, this embodiment also provides an automatic target detection system for SAR images, including a microprocessor and a memory connected to each other, and the microprocessor is programmed or configured to execute the steps of the aforementioned automatic target detection method for SAR images. In addition, this embodiment also provides a computer-readable storage medium, where a computer program programmed or configured to execute the aforementioned method for automatic target detection oriented to SAR images is stored.

Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems, or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein. The present application is described with reference to flowcharts and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each procedure and/or block in the flowchart and/or block diagram, and a combination of procedures and/or blocks in the flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a An apparatus for realizing the functions specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram. These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in one or more procedures of the flowchart and/or one or more blocks of the block diagram. These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in the flow chart or blocks of the flowchart and/or the block or blocks of the block diagrams.

The above descriptions are only preferred implementations of the present invention, and the protection scope of the present invention is not limited to the above-mentioned embodiments, and all technical solutions under the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for those skilled in the art, some improvements and modifications without departing from the principles of the present invention should also be regarded as the protection scope of the present invention.

Claims (9)

1. A target automatic detection method facing SAR images is characterized by comprising the steps of inputting input images into a feature fusion and anti-interference network FFAN which is trained in advance to obtain a target frame of target detection, wherein the feature fusion and anti-interference network FFAN comprises the following steps:

a backbone network, which is used for down-sampling the input image into features of different scales, including low-level features P3, middle-level features P4-P6 and high-level features P7;

the pooling and channel shuffling module is used for extracting and fusing the intermediate-level features P4-P6 respectively to obtain corresponding shuffling features P4 '-P6';

the depth fusion module is used for performing depth fusion on the low-level features P3, the shuffle features P4 'to P6' and the high-level features P7 to respectively obtain the corresponding low-level features P3', the shuffle features P4' to P6 'and the high-level features P7' after depth fusion;

the spatial attention mechanism RSAM is respectively used for further spatial feature extraction and fusion of the deeply fused shuffle feature P6 'and the high-level feature P7';

the spatial feature fusion module ASFF is used for further spatial feature extraction and fusion of the deeply fused low-level features P3' and the shuffled features P4' to P5' respectively;

the classification regression network is used for performing classification regression on all spatial features output by the spatial attention mechanism RSAM and the spatial feature fusion module ASFF through extraction and fusion to generate a target frame for target detection;

the pooling and channel shuffling module comprises three shuffling modules including a shuffling module PSM-l, a shuffling module PSM-m and a shuffling module PSM-s, wherein the shuffling module PSM-l is used for extracting and fusing intermediate-level features P4 to obtain corresponding shuffling features P4 respectively, the shuffling module PSM-m is used for extracting and fusing intermediate-level features P5 to obtain corresponding shuffling features P5 respectively, the shuffling module PSM-s is used for extracting and fusing intermediate-level features P6 to obtain corresponding shuffling features P6 respectively, and the shuffling module PSM-l, the shuffling module PSM-m and the shuffling module PSM-s respectively comprise:

the channel separation module is used for dividing the channel of the input characteristics into two parts;

the maximum pooling operation module is used for performing maximum pooling operation on a part of channels output by the channel separation module;

the average pooling operation module is used for carrying out average pooling operation on the other part of channels output by the channel separation module;

the splicing module is used for splicing the output of the maximum pooling operation module, the output of the average pooling operation module and the original input characteristics of the input channel separation module;

a shuffling module for shuffling the output of the splicing module to enrich the information of each channel;

the convolution module is used for carrying out convolution operation on the output of the shuffling module to enable the number of output characteristic channels to be the same as that of input characteristics;

the sizes of the pooling cores of the shuffling module PSM-l, the shuffling module PSM-m and the shuffling module PSM-s for performing the maximum pooling operation and the average pooling operation are 9,15,25, 5,11 and 3 respectively.

2. The SAR image-oriented target automatic detection method according to claim 1, wherein the backbone network is an Efficientnet network, the Efficientnet network sequentially performs down-sampling of different sizes of 512 × 512 input images respectively to obtain corresponding features P1 to P7, and the features obtained by the 3 rd to 7 th down-sampling are respectively used as a low-level feature P3, middle-level features P4 to P6 and a high-level feature P7.

3. The SAR image-oriented target automatic detection method according to claim 1, characterized in that the depth fusion module is an improved structure of BiFPN, and the improved structure adds four jump connections on the basis of BiFPN, wherein the first jump connects the middle-level feature P6 and the shuffle feature P5', the second jump connects the middle-level feature P5 and the shuffle feature P4', the third jump connects the middle-level feature P5 and the middle-level feature P7, and the fourth jump connects the middle-level feature P4 and the middle-level feature P6.

4. The SAR image-oriented target automatic detection method according to claim 1, characterized in that the classification regression network comprises a classification network and a regression network, the classification network adopts 3 times of convolution of 64 channels and 1 time of convolution of n _ b × n _ c to predict the category of each target frame, the regression network adopts 3 times of convolution of 64 channels and 1 time of convolution of n _ b × 4 to predict the regression condition of each prior frame, wherein n _ b refers to the number of prior frames owned by the feature layer, and n _ c refers to the number of network target detection categories.

5. The SAR image-oriented target automatic detection method according to claim 1, wherein a step of pre-training a feature fusion and anti-interference network FFAN is further included before inputting an input image into a feature fusion and anti-interference network FFAN which is trained in advance to obtain a target frame for target detection, and a loss function L adopted when training the feature fusion and anti-interference network FFAN _AIF Comprises the following steps:

L _AIF ＝L _FL +L _AIoU ，

in the above formula, L _FL Is the Focal loss classification loss, L _AIoU Is AIoU regression loss, and has:

L _AIoU ＝L _CIoU +pw，

in the above formula, L _CIoU For CIoU regression loss, a and a ^t Respectively representing the areas of the prediction frame and the real frame, wherein p is a weight function of the areas, w is the similarity of the areas, and IoU is an intersection ratio.

6. An automatic target detection method for SAR images is characterized by comprising the following steps:

s1) generating a window image from an input SAR image in a sliding window mode with an overlapping area;

s2) taking each window image as an input image, and carrying out target detection by adopting the steps of the SAR image-oriented target automatic detection method in any one of claims 1-5 to obtain a target frame;

s3) converting the target frame coordinates of each window image into the original SAR image;

and S4) filtering the target frames overlapped in the original SAR image to obtain the target frame serving as the final target detection result.

7. The SAR image-oriented target automatic detection method as claimed in claim 6, characterized in that the method adopted in step S4) for filtering the target frames overlapped in the original SAR image is a non-maximum suppression method.

8. An automatic target detection system for SAR images, comprising a microprocessor and a memory connected to each other, characterized in that the microprocessor is programmed or configured to execute the steps of the automatic target detection method for SAR images according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that the computer-readable storage medium has stored therein a computer program programmed or configured to execute the SAR image-oriented object automatic detection method according to any one of claims 1 to 7.

Images