A SAR ship detection method based on receptive field weight regularization

Abstract

The application discloses a SAR ship detection method based on a receptive field weight regularization, and relates to the technical field of SAR target detection.The detection method comprises the following steps: extracting spatial features based on multiple receptive fields, and reducing the coupling between different features through a regularization weight;enhancing the expression ability of ship features by combining the information of surrounding objects through a context enhancement module; and performing classification and positioning tasks respectively by a CIoU-based detection head.The detection method can extract rich spatial features, reduce the coupling between different features by using a regularization weight, and enhance the expression ability of ship features by using the context enhancement module, so that the distinguishability of ship features is improved, and the CIoU-based detection head can obtain more accurate prediction boxes.Compared with the current mainstream target detection methods, the detection accuracy is high.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of SAR target detection, and particularly relates to a SAR ship detection method based on receptive field weight regularization. BACKGROUND

[0002] Synthetic aperture radar (SAR) technology is an active microwave remote sensing technology, which uses pulse compression and synthetic aperture to improve the range and azimuth resolution of the radar, and obtains a high-resolution, large-area SAR image in all-weather and all-day. SAR image recognition technology is one of the key technologies in artificial image interpretation, and is mainly used to avoid the influence of inherent noise, obtain potential feature information of targets in a region of interest (RoI), and provide strong data support for target detection. Ships are important carriers for marine transportation, and ship detection is an important part of marine surveillance. Based on the characteristics of SAR images and the importance of ship detection, research on ship target detection based on SAR images has received widespread attention in the fields of military reconnaissance, marine resource monitoring and marine transportation management.

[0003] Target detection, as one of the research directions of computer vision, can provide valuable information for semantic understanding of images and videos. Its essence is to locate and classify objects, and the main task is to accurately and effectively find all objects of interest in a given image. According to whether manual feature extraction is required, target detection algorithms can be divided into traditional methods and deep learning-based methods. Traditional target detection methods can be divided into three categories: 1) statistical-based target detection; 2) knowledge-based target detection; and 3) model-based target detection. Generally, these traditional methods include three steps: region selection, feature extraction and classification. The region selection part is completed using an exhaustive strategy, which results in a high time complexity of the method. Then, in the feature extraction aspect, hand-crafted methods such as scale-invariant feature transform, histogram of oriented gradients and speeded-up robust features are used. However, the diversity of object shapes, the diversity of light changes and the diversity of backgrounds in images will make the extracted features less robust. Therefore, the recognition effect and accuracy of traditional methods are not high, and they will also take a lot of time.

[0004] With the rapid development of deep learning theory, the target detection method based on deep learning shows better performance than traditional methods. At present, the target detection method based on deep learning mainly includes two types. The first type is a two-stage detection method, also known as a region-based method. Fast-RCNN and Faster-RCNN are representative methods in this type of method. In this type of method, basic features are extracted through a backbone network first, and then the features are segmented according to a pre-set anchor box to obtain regions of interest (RoI). The features in the RoI are reconstructed to a fixed size, and then processed through a classification and regression branch. Finally, the processed features are input to perform non-maximum suppression (NMS) to obtain the detection result. The other type is a one-stage detection method. The network directly predicts the position and classification label of a potential target on multiple feature maps without remodeling and adjusting the RoI. Representative methods include YOLO and SSD. The structure of these methods is simpler than that of the two-stage method, and the detection speed is also faster than that of the two-stage method. With the improvement of the target detection framework, the detection accuracy of the one-stage detector can be comparable to that of the two-stage detector. Therefore, more and more scholars are studying the one-stage detector and combining the one-stage detector with the two-stage detector. These algorithms have been widely applied in the field of ship detection and identification.

[0005] The prior art has the following disadvantages:

[0006] Although the performance of the target detection method has been greatly improved, most of these methods are aimed at natural images and do not take into account the inherent characteristics of SAR images. The ship features extracted by the existing methods are not sufficient, which leads to a low recognition rate of ships in SAR images. In the near-shore scene, some interference exhibits similar features to the target, which increases the difficulty of correctly judging the target. SUMMARY

[0007] The purpose of the present application is to provide a SAR ship detection method based on receptive field weight regularization. First, rich spatial features are extracted using multiple receptive fields, and the coupling between different features is reduced using regularization weights. Second, the proposed context enhancement module can combine the information of surrounding objects to enhance the expression ability of ship features and improve the distinguishability of ship features. Third, different branches are used to perform classification and regression tasks respectively, avoiding information loss. The CIoU-based regression loss is beneficial to obtain more accurate prediction boxes to solve the deficiencies in the background art.

[0008] In order to achieve the above purpose, the present application provides the following technical scheme: a SAR ship detection method based on receptive field weight regularization, the detection method comprising the following steps:

[0009] S1: extract spatial features based on multi-receptive field, and reduce the coupling between different features through regularizing weights;

[0010] S2: enhance the expression ability of ship features by combining the information of surrounding objects through a context enhancement module;

[0011] S3: the detection head based on CIoU performs classification and positioning tasks respectively, and uses the extracted feature information to guide the prediction of the regression branch through the CIoU loss function.

[0012] In a preferred embodiment, in step S1, the spatial feature extraction is based on the RFWR module, which includes five branches:

[0013] The first branch is composed of two 1x1 convolution filters, and the feature F1 is the output of this branch;

[0014] The second branch is obtained by 1x1 filter and 3x3 filter convolution, and the feature F2 is obtained;

[0015] The third branch is obtained by 1x1 filter and 3x3 filter convolution, and the 3x3 filter in the third branch is an expansion convolution filter with an expansion rate of 3, and the output is F3;

[0016] The fourth branch is composed of a 1x1 convolution filter and a 5x5 expansion convolution, with an expansion rate of 5 and an output of F4;

[0017] The fifth branch is a weight regularization, composed of a global average pooling layer and a Softmax layer, which regularizes the weights by introducing Softmax, and fuses different spatial features by weighting, so that the fused features include global features and different local features.

[0018] In a preferred embodiment, the context enhancement module includes a CEM module, which is used to extract context information.

[0019] In a preferred embodiment, the CEM module includes a global pooling branch and two parallel convolution branches;

[0020] The parallel convolution branch is used to aggregate context information of multiple receptive fields with different sizes, wherein one parallel branch contains a 3x3 convolution filter, and the other parallel branch contains two 3x3 convolution filters, and the output features of different branches are spliced through Concat operation;

[0021] The global pool branch is used to introduce global information, the global context information is captured by the global pooling layer, processed by 1x1 convolution, and the final output of the global pool branch is obtained, the final output value is connected between the parallel convolution branch and the global pool branch, and 1x1 convolution processing is performed on the connected feature map to obtain the final output with the same shape as the input.

[0022] In a preferred embodiment, the CEM module obtains the final output with the same shape as the input, and the calculation formula is:

[0023] c1 = Conv 3x3 (input)

[0024] c2 = Conv 3×3 (Conv 3×3 (input))

[0025] g = Conv 1×1 (G(input))

[0026] Fuse = Conv 1×1 (Concat(Concat(c1, c2), g))

[0027] In the formula, input represents input features, Conv 1x1 () and Conv 3x3 () represent 1x1 and 3x3 convolution respectively, c1, c2 and g are the outputs of the first, second and global pool branches respectively, G() and Concat() represent global pooling operation and feature concatenation operation, and Fuse is the final output of CEM.

[0028] In a preferred embodiment, the CIoU-based detection head is composed of three branches of classification, target and regression, the classification branch is used to generate confidence score of the category contained by each feature point, the regression parameter is obtained by the regression branch, the prediction box is generated by adjusting the regression parameter, and the target branch is used to judge whether each feature point contains an object.

[0029] In a preferred embodiment, the calculation expression of the CIoU loss function is:

[0030] L = λ1L cla + λ2L reg + λ3L obj

[0031] In the formula, L is the CIoU loss function, L cla , L reg and L obj are the loss functions of the classification branch, the regression branch and the target branch respectively, and λ1, λ2 and λ3 represent their respective weights.

[0032] In a preferred embodiment, the calculation expression of the regression branch loss function is:

[0033]

[0034] In the formula, L reg is the loss function of the regression branch, d represents the distance between the center point of the prediction box and the ground truth box, and C represents the diagonal distance of the minimum closed region containing the prediction true value box and the ground truth box at the same time.

[0035] The calculation expressions of alpha and v are:

[0036]

[0037] In the formula, w gt and h gt represent the width and height of the ground truth box, and w p and h p are the width and height of the prediction box.

[0038] In the above technical solution, the technical effects and advantages provided by the present application are:

[0039] 1. The detection method can extract rich spatial features and reduce the coupling between different features by using regularization weights. In addition, the proposed context enhancement module can combine the information of surrounding objects to enhance the expression ability of ship features and improve the distinguishability of ship features. The detection head based on CIoU can obtain more accurate prediction boxes. Compared with the current mainstream target detection method, the SAR ship detection method based on receptive field weight regularization and context enhancement has high detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments or prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0041] Figure 1 It is a network structure schematic diagram of the present application.

[0042] Figure 2 It is a structure schematic diagram of the RFWR module of the present application.

[0043] Figure 3 It is a context enhancement module schematic diagram of the present application.

[0044] Figure 4The schematic diagram of the decoupling head with CIoU of the application.

[0045] Figure 5 The PR curve schematic diagram of the ablation experiment of the application.

[0046] Figure 6 The PR curve schematic diagram of the detection results of different methods of the application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0048] Embodiment 1: please refer to Figure 1 As shown in the figure, the SAR ship detection method based on receptive field weight regularization described in the embodiment includes the following steps:

[0049] Rich spatial features are extracted by using multi-receptive field, and the coupling between different features is reduced by using regularization weight. The feature extraction combined with receptive field weight regularization can extract receptive field features of different sizes, improving the feature extraction capability of the method.

[0050] The context enhancement module combines the information of surrounding objects to enhance the expression ability of ship features and improve the distinguishability of ship features. The context module fuses the background information of the SAR ship, enhances the representation ability of the features, and helps accurate detection of the target.

[0051] The detection head based on CIoU adopts a decoupling detection head to perform classification and positioning tasks respectively, fully utilizes the extracted feature information, uses the CIoU loss function to guide the prediction of the regression branch, makes the predicted box more consistent with the ground true box, and improves the accuracy of the method.

[0052] The application can extract rich spatial features and reduce the coupling between different features by using regularization weight. Secondly, the proposed context enhancement module can combine the information of surrounding objects to enhance the expression ability of ship features and improve the distinguishability of ship features. The detection head based on CIoU can obtain more accurate prediction box. Compared with the current mainstream target detection method, the SAR ship detection method based on receptive field weight regularization and context enhancement has high detection accuracy.

[0053] Embodiment 2: The feature extraction combined with receptive field weight regularization can extract features of different sizes of receptive fields, and improve the feature extraction capability of the method.

[0054] Feature extraction maps the sample data set from a high-dimensional feature space to a low-dimensional feature space, and the sample data set after mapping still maintains good distinguishability, and the performance of feature extraction directly affects the final recognition result of the target, and features with high recognition can improve the probability of correct detection of objects, so it is of great significance to extract features with high recognition for target detection, and on the basis of the combination of receptive field size and eccentricity, the application increases the size of the receptive field by designing a large dilated rate convolution filter to enhance the features and highlight the center information of the ship target.

[0055] In order to extract more useful ship features, the application proposes an RFWR module to improve the spatial feature extraction capability of the network, and by designing multiple dilated convolution branches with different expansion rates, the size of the receptive field can be enlarged, and the central feature information can be highlighted, and the specific structure of the RFWR is as shown in Figure 2 .

[0056] As shown in Figure 2 , the RFWR has five branches, the first branch is composed of two 1x1 convolution filters, and the feature F1 is the output of the branch, in the second branch, F2 is obtained after convolution by 1x1 filter and 3x3 filter, the third branch is similar to the second branch, but the 3x3 filter in the third branch is a dilated convolution filter with an expansion rate of 3, and the output is represented by F3, and F4 is obtained from the fourth branch, which is composed of 1x1 convolution filter and 5x5 dilated convolution, and the expansion rate is 5, in order to reduce the coupling between features, the fifth branch is designed as weight regularization, which is composed of global average pooling layer and Softmax layer, the weight is regularized by introducing Softmax, and different spatial features are fused by weighting, so that the fused features include global features and different local features.

[0057] Embodiment 2: The context enhancement module combines the information of surrounding objects to enhance the expression capability of ship features and improve the distinguishability of ship features, the context module fuses the background information of SAR ship, enhances the representation capability of features, and is helpful for accurate detection of targets.

[0058] Contextual feature enhancement is designed to better utilize the extracted feature information, which includes CBS, UpSampling, CCSP and CEM. Contextual information is very important in object detection tasks, which describes the relationship between object location and its surrounding area. In SAR images with complex background, the surrounding area can provide useful information for target detection, which is beneficial to detect targets with unobvious features. Therefore, the context considers the semantic and location information of the surrounding area of the target, combines the features of the surrounding area of the ship from a global perspective, and improves the recognizability of the target of interest and the accuracy of the detection method.

[0059] As shown in Figure 3 , CEM is used to extract contextual information, which has two parallel convolution branches and one global pooling branch. The parallel convolution branch is used to aggregate contextual information with multiple receptive fields of different sizes. One parallel branch only contains a 3x3 convolution filter, and the other contains two 3x3 convolution filters. Then the output features of different branches are spliced through Concat operation. The global pool branch is used to introduce global information. The global contextual information can be captured by the global pooling layer, and then processed by 1x1 convolution to obtain the final output of this branch. The final output values are connected between the parallel convolution branch and the global pooling branch, and the connected feature maps are processed by 1x1 convolution to obtain the final output with the same shape as the input. The calculation formula is:

[0060] c1=Conv 3x3 (input)

[0061] c2=Conv 3×3 (Conv 3×3 (input))

[0062] g=Conv 1×1 (G(input))

[0063] Fuse=Conv 1×1 (Concat(Concat(c1,c2),g))

[0064] In the formula, input represents the input feature, Conv 1x1 () and Conv 3x3 () represent 1x1 and 3x3 convolution respectively, c1, c2 and g are the outputs of the first, second and global pooling branches respectively, G() and Concat() represent the global pooling operation and feature concatenation operation, and Fuse is the final output of CEM. This module can fuse the contextual information of the object in the feature, enhance the feature, and improve the accuracy of object detection.

[0065] Embodiment 3: The CIoU-based detection head uses a decoupled detection head to perform classification and positioning tasks respectively, fully utilizes the extracted feature information, and uses the CIoU loss function to guide the prediction of the regression branch, so that the predicted box is more consistent with the ground truth box, and the accuracy of the method is improved.

[0066] As shown in Figure 4 , after the backbone extracts the feature information of the image, the detection head obtains the classification, position and size of the object according to the extracted information. In target detection, the detection head needs to complete two tasks of classification and positioning. Studies have found that the focal areas of the two tasks are different. When a set of parameters is shared by the classification task and the positioning task, information loss will occur. Considering this, the application uses a decoupled head to perform classification and positioning tasks using different branches. Based on this head, multiple predictions of different objects at different positions can be flexibly collected and aggregated.

[0067] The CIoU-based detection head is composed of three branches of classification, target and regression. The classification branch is mainly used to generate the confidence score of the category contained by each feature point. The regression parameter is obtained by the Reg branch, and the prediction box is generated by adjusting the regression parameter. The target branch is used to judge whether each feature point contains an object.

[0068] The loss function is used to measure the difference between the predicted data and the ground truth. In network training, the difference between the forward calculation result and the ground truth is calculated to guide the direction of subsequent training. In this head, cross-entropy is used as the loss function of the classification and target branches, and the calculation expression is:

[0069] Loss(y i gt ,y i p )=-(y i gt log(y i p )+(1-y i gt )log(1-y i p ))

[0070] where y i gt represents the true value, and y i p is the predicted value. The regression branch uses CIoU as the loss function to complete the training. CIoU is an improved version of IoU, which can be used to represent the degree of overlap between the candidate box and the ground truth. Then, a prediction box is selected from the candidate box.

[0071] The calculation expression of IoU is:

[0072]

[0073] In the formula, B gt and B p represent ground-truth and predicted box respectively, but IoU only reflects the size of the overlapping area of them, and cannot reflect the overlapping manner of these boxes, and more importantly, when there is no overlapping area between these boxes, IoU loses its meaning, so not only the overlapping area should be considered, but also the position of the center point and the aspect ratio of the detection box.

[0074] Therefore, the present application adopts CIoU as the loss function of the regression branch, and the expression is as follows:

[0075]

[0076] In the formula, L reg is the loss function of the regression branch, d represents the distance between the center point of the predicted box and the ground truth box, and C represents the diagonal distance of the smallest closed area containing the predicted true value box and the ground truth box.

[0077] The calculation expressions of α and v are as follows:

[0078]

[0079]

[0080] In the formula, w gt and h gt represent the width and height of the ground truth box, and w p and h p are the width and height of the predicted box.

[0081] The CIoU loss function is the weighted sum of the three loss functions, and the calculation expression is as follows:

[0082] L = λ1L cla + λ2L reg + λ3L obj

[0083] In the formula, L is the CIoU loss function, L cla , L reg and L obj are the loss functions of the classification branch, the regression branch and the target branch respectively, and λ1, λ2 and λ3 represent their respective weights.

[0084] Embodiment 4: In this embodiment, the above detection method is experimented, and the specific process is as follows:

[0085] I. Experimental data and experiments

[0086] The environmental platform adopts Intel(R) Xeon(R) Silver 4210 CPU@2.20GHz, 32gb memory, NVIDIA Quadro RTX5000, 64-bit windows 10 operating system, the programming language used is python, the integrated development environment is Pycharm, Tensorflow-gpu, opencv, PIL, keras and numpy are the packages used in the development process, the optimization process is completed by Adam, GPU accelerated training, all parameters are shown in table 1:

[0087] Table 1 parameter setting table

[0088]

[0089] The SAR images in SSDD have different resolutions, sensors, polarizations, sea conditions and scenes, the scenes are divided into nearshore and offshore, the size of the ships in different scenes is different, and the data diversity is the basis for establishing a reliable detection model, therefore, SSDD is selected as the data set, the data set is constructed according to the annotation format of PASCAL VOC, and is composed of 1160 images, in order to compare fairly, the data set is divided into a training set and a test set according to a ratio of 4:1, and all methods are trained and tested on the training set and the test set.

[0090] II. Evaluation index

[0091] In order to quantitatively analyze the performance of the algorithm, precision, recall, F1 and AP are used as evaluation indexes, the precision is the ratio of the targets correctly detected by the detector to all the detected targets, the recall is the ratio of the correctly detected targets to all the ground truth targets, they are commonly used evaluation indexes, and the definition formulas are as follows:

[0092]

[0093]

[0094] Among them, true positive (TP) is the number of targets correctly detected by the detector, false positive (FP) represents the number of false alarm targets, false negative (FN) is the number of targets missed by the detector, F1 is the overall measure of model precision, which is the combination of precision and recall, it can better evaluate the accuracy and recall of the model, F1 score is good, which means that the P and R of the detector are low, therefore, as many targets as possible can be correctly identified, and the interference of false positives is minimized, which can be expressed by the following formula:

[0095]

[0096] By setting different R, the corresponding P is obtained, and then taking R as the horizontal axis and P as the vertical axis, the PR curve is obtained, and the area covered by the PR curve is AP, which can be expressed as:

[0097] III. Ablation experiment

[0098] The present application respectively proposes RFWR and CEM to improve the feature extraction capability and combine the context information. In order to prove the effectiveness of these modules, an ablation experiment is designed to analyze the influence of the two modules on the overall detection performance of the model. The detector without RFWR and CEM is taken as the baseline, and then the two modules are added to the baseline in turn for experiment. By comparing the four evaluation indexes, the detection results of different detectors are evaluated, and the influence of the two proposed modules is analyzed. In order to ensure the fairness of comparison, the same parameters are used in the ablation experiment, as shown in Table 2:

[0099] Table 2 Ablation experiment results

[0100]

[0101] Table 2 shows the ablation results of WRFB and CEM modules. The baseline is the detector without WRFB and CEM modules, and its P, R, F1 and AP0.5 are 92.98%, 91.04%, 0.9200 and 90.12% respectively. After embedding the CEM module based on the baseline, the evaluation indexes are improved, and P, R, F1 and AP0.5 are increased by 3.59%, 2.71%, 0.0314 and 2.85% respectively. This is because combining context information can improve the recognizability of objects and the possibility of correctly detecting objects. Embedding WRFB into the Baseline+CEM model, R and AP0.5 are increased by 3.54% and 3.22% respectively. The reason is that the WRFB module can extract features of different receiving fields, enhance the spatial feature extraction capability of the network, and make the object easy to identify. Figure 5 The PR curve corresponding to Table 2 further shows that CEM and WRFB can improve the detection performance and prove its effectiveness.

[0102] IV. Comparison with other methods

[0103] Table 3 Comparison of detection results of different methods

[0104]

[0105] Table 3 shows the detection results of different methods. The R, F1 and AP0.5 of the method in this paper achieve the best results. The PR curves of the detection results of different methods are as follows: Figure 6As shown, it can be clearly seen from the figure that the area covered by the light blue line is the smallest, the area covered by the purple line is slightly higher than that of the dark blue line, the area covered by the green line is the fourth largest line, the areas covered by the pink and yellow lines are the same, and the area covered by the red line is the largest. These results are consistent with the data shown in Table 3, further verifying the effectiveness of the method.

[0106] The above embodiments can be implemented wholly or partially by software, hardware, firmware, or any combination thereof. When implemented by software, the above embodiments can be implemented wholly or partially in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through a wired or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0107] It should be understood that the term "and / or" in the present application only describes the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, A and B exist together, and B exists alone, where A and B can be singular or plural. In addition, the character " / " in the present application generally represents an "or" relationship between the associated objects before and after it, but it can also represent an "and / or" relationship, which can be understood in the context before and after it.

[0108] In the present application, "at least one" means one or more, and "multiple" means two or more. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b, or c can mean a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, and c can be single or multiple.

[0109] It should be understood that the size of the sequence number of each process described above does not mean the order of execution, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0110] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present application can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0111] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0112] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0113] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiments.

[0114] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit.

[0115] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0116] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A SAR ship detection method based on receptive field weight regularization, characterized in that: The detection method includes the following steps: S1: Spatial features are extracted based on multiple receptive fields, and the coupling between different features is reduced by regularizing the weights; spatial feature extraction is implemented based on the RFWR module, which includes five branches: The first branch consists of two 1×1 convolutional filters, and feature F1 is the output of this branch; The second branch is convolved with a 1×1 filter and a 3×3 filter to obtain F2; The third branch is obtained by convolving a 1×1 filter and a 3×3 filter. The 3×3 filter in the third branch is a dilated convolution filter with a dilation rate of 3, and the output is F3. The fourth branch consists of a 1×1 convolutional filter and a 5×5 dilated convolution, with a dilation rate of 5, resulting in an output F4. The fifth branch is weight regularization, which consists of a global average pooling layer and a Softmax layer. By introducing Softmax, the weights are regularized, and different spatial features are fused by weighted fusion, so that the fused features include global features and different local features. S2: Enhance the expressive power of ship features by combining information from surrounding objects through the context enhancement module; the context enhancement module includes a CEM module, which is used to extract contextual information; S3: The CIoU-based detection head performs classification and localization tasks respectively. It uses the extracted feature information and guides the prediction of the regression branch through the CIoU loss function. The CIoU-based detection head consists of three branches: classification, target, and regression. The classification branch is used to generate the confidence score of the category contained in each feature point. The regression parameters are obtained by the regression branch. The prediction box is generated by adjusting the regression parameters. The target branch is used to determine whether each feature point contains an object.

2. The SAR ship detection method based on receptive field weight regularization according to claim 1, characterized in that: The CEM module includes a global pooling branch and two parallel convolutional branches; Parallel convolutional branches are used to aggregate contextual information from multiple accepting fields of different sizes. One parallel branch contains a 3×3 convolutional filter, and another parallel branch contains two 3×3 convolutional filters. The output features of different branches are concatenated through the Concat operation. The global pooling branch is used to introduce global information. Global context information is captured through the global pooling layer and processed by 1×1 convolution to obtain the final output of the global pooling branch. The final output value is connected between the parallel convolution branch and the global pooling branch, and the connected feature map is processed by 1×1 convolution to obtain the final output with the same shape as the input.

3. The SAR ship detection method based on receptive field weight regularization according to claim 2, characterized in that: The CEM module obtains a final output with the same shape as the input, calculated using the following formula: ； ； ； ； In the formula, input represents the input feature, and Conv 1x1 () and Conv 3x3 () represent 1×1 and 3×3 convolutions respectively, c1, c2 and g are the outputs of the first, second and global pooling branches respectively, G() and Concat() represent the global pooling operation and feature concatenation operation respectively, and Fuse is the final output of CEM.

4. The SAR ship detection method based on receptive field weight regularization according to claim 1, characterized in that: The calculation expression for the CIoU loss function is as follows: ； In the formula, L is the CIoU loss function, L cla L reg and L obj λ1, λ2, and λ3 are the loss functions for the classification branch, regression branch, and target branch, respectively, and represent their respective weights.

5. The SAR ship detection method based on receptive field weight regularization according to claim 4, characterized in that: The calculation expression for the regression branch loss function is as follows: ； In the formula, Here, d represents the distance between the center point of the predicted box and the ground truth box, and C represents the diagonal distance of the smallest closed region that simultaneously contains both the predicted and ground truth boxes. The expressions for calculating α and v are: ； ； In the formula, w gt and h gt w represents the width and height of the ground plane's actual bounding box. p and h p These are the width and height of the prediction box.

Images