SAR ship directional detection method, device, equipment and medium based on deep learning

Abstract

The present application relates to a SAR ship directional detection method, device, equipment and medium based on deep learning, which extracts features from SAR measured images by using a multi-scale large convolution kernel feature extraction network to obtain a multi-scale feature map, the network includes multiple multi-scale large convolution kernel feature extraction layers connected in sequence, and uses heterogeneous deep convolution to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer, and uses a feature fusion network including multiple feature fusion layers to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, each feature fusion layer dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level and low-level scale feature maps of the corresponding level to generate a distinguishable multi-scale pyramid feature map, and finally the detection head outputs a ship target directional detection frame according to the multi-scale pyramid feature map. The present method can effectively improve the accuracy and efficiency of ship target detection.

Description

Technical Field

[0001] The present application relates to the technical field of SAR target detection, and in particular to a SAR ship directional detection method, device, equipment and medium based on deep learning. Background Art

[0002] Synthetic aperture radar (SAR) is a high-resolution imaging radar technology that uses radar platforms such as aircraft and satellites to transmit and receive radio waves in motion to obtain high-resolution images of ground or sea targets. Due to its unique imaging mechanism, SAR can independently and continuously observe the ground. In addition, SAR also has strong penetration capability and high-resolution imaging capabilities. Therefore, SAR has important applications in disaster management, resource exploration, urban planning, ocean monitoring and other fields.

[0003] With the continuous development of SAR imaging technology, SAR images have been widely used for the detection and identification of ship targets, which is of great significance to maritime safety and rights protection. Traditional SAR ship detection technology mainly relies on signal processing and image analysis methods, including but not limited to edge detection, threshold segmentation, superpixel detectors, etc. Although traditional SAR ship target detection technology has promoted the development of marine target detection to a certain extent, it also has obvious limitations and challenges. First, these traditional methods have poor environmental adaptability and limited ability to suppress interference such as sea clutter, shadows, and sidelobe echoes. Secondly, the selection of these parameters is highly dependent on the experience of the operator and may need to be adjusted frequently in different situations. In addition, the feature expression ability of traditional methods is limited. Compared with the currently popular deep learning (DL) methods, these traditional methods have a low degree of automation when processing large-scale data sets and require more manual intervention. Unlike traditional detection methods, DL-based methods can cope with complex environments or backgrounds more effectively. At the same time, it overcomes the time-consuming and labor-intensive shortcomings of feature design in traditional image interpretation, has more advantages in target detection and recognition, and greatly promotes the improvement of target detection technology.

[0004] However, despite some progress, DL-based SAR ship detection methods for arbitrary directions still face some challenges. First, traditional CNNs lack the global modeling capabilities required for complex SAR images, making it difficult to effectively characterize multi-scale ship targets in complex scenes. Unlike optical images, SAR images themselves contain less information, and the extensive use of feature extraction structures exacerbates the problem of target scale imbalance. In addition, in terms of feature utilization, current SAR ship detection methods rely on the pyramid structure in general target detection algorithms for feature fusion, ignoring the different contributions of spatial and channel features in SAR images in the feature fusion process. This phenomenon is considered to be feature utilization imbalance in SAR ship detection. In addition, the current SAR ship detection strategy adapted to arbitrary directions adds an angle branch to the model head to calculate the IoU distance loss for angle regression and combines it with the classification loss of multi-task feedback. However, this method does not address the inherent feature learning limitations, is not suitable for SAR ship targets, and leads to significant regression loss imbalance. Summary of the invention

[0005] Based on this, it is necessary to provide a deep learning-based SAR ship directional detection method, device, equipment and medium that can improve detection accuracy and efficiency in response to the above technical problems.

[0006] A SAR ship directional detection method based on deep learning, the method comprising:

[0007] Acquire SAR measured images with ships as targets;

[0008] A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0009] A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0010] A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

[0011] In one embodiment, in the multi-scale large convolution kernel feature extraction network, the SAR measured image is first processed through the starting layer to obtain an initial feature map, and then the initial feature map is input into the multi-scale large convolution kernel feature extraction layer;

[0012] A partitioned convolution method is used in each of the multi-scale large convolution kernel feature extraction layers to send the input feature map to the upper branch and the lower branch respectively;

[0013] In the lower branch, the input feature map is processed by convolution and then passed through a large convolution kernel feature extraction block for feature extraction. The extracted feature map is fused with the feature map passed through the upper branch, and then the output feature map is obtained through the attention channel and the convolution layer. The convolution kernel size of the large convolution kernel feature extraction block is determined by the number of layers in the multi-scale large convolution kernel feature extraction network.

[0014] Among them, in the large convolution kernel feature extraction block, the feature map after convolution processing is feature grouped, and then the grouped feature map is processed by using the reverse bottleneck layer to obtain the output feature map of the lower branch.

[0015] In one embodiment, in the large convolution kernel feature extraction block:

[0016] The grouped feature maps are processed using an iterative processing method, and the first group of feature maps remain unchanged;

[0017] The first set of feature maps and the second set of feature maps are fused, and then a reverse bottleneck layer is used to extract features to obtain a first fused feature map;

[0018] The first fused feature map and the third set of feature maps are fused, and then a reverse bottleneck layer is used to perform feature extraction to obtain a second fused feature map;

[0019] The first group of feature maps, the first fused feature map and the second fused feature map are combined along the channel axis and input into the channel attention mechanism for dynamic adjustment at the channel level to obtain input features of the large convolution kernel feature extraction block.

[0020] In one embodiment, the feature maps of different scales in the first three layers obtained by the multi-scale large convolution kernel feature extraction network are input into the feature fusion network for feature fusion enhancement.

[0021] In one embodiment, in each of the feature fusion layers:

[0022] Adjusting the spatial and channel scales of the high-level scale feature map and the low-level scale feature map to obtain a high-level scale input feature map and a low-level scale input feature map;

[0023] In the global channel feature optimization stage, the high-level scale input feature map and the low-level scale input feature map are fused, and a global channel attention weight is obtained based on the fused feature;

[0024] In the global spatial feature optimization stage, convolution is used to process the high-level scale input feature map and the low-level scale input feature map to adjust the corresponding weights, and the two weights are added element by element to obtain the global spatial attention weight;

[0025] A corresponding scale pyramid feature map output by the feature fusion layer is obtained according to the global channel attention weight, the global spatial attention weight and the fusion feature, wherein the fusion feature is a feature obtained by fusing the high-level scale input feature map and the low-level scale input feature map.

[0026] In one embodiment, a ship target directional detection network is constructed according to the multi-scale large convolution kernel feature extraction network, the feature fusion network and the detection head, and the SAR measured image is input into the ship target directional detection network for target directional detection. When constructing a training data set for the ship target directional detection network:

[0027] Acquire a data set with rotation bounding box annotations, wherein the data set includes a plurality of SAR images with a ship as a target;

[0028] The data set is expanded to obtain an enhanced data set, and the enhanced data set is used as the training data set.

[0029] In one embodiment, when training the ship target directed detection network, boundary regression loss and classification loss are used for training, wherein the boundary regression loss is represented by optimizing the loss by combining the distance constraint between Gaussian distributions.

[0030] The present application also provides a SAR ship directional detection device based on deep learning, the device comprising:

[0031] SAR measured image acquisition module, used to acquire SAR measured images with ships as targets;

[0032] A multi-scale feature extraction module, used to extract features from the SAR measured image using a multi-scale large convolution kernel feature extraction network to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0033] A multi-scale fusion feature extraction module, configured to use a feature fusion network to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, each of which dynamically optimizes channel features to generate the multi-scale pyramid feature map with discriminability by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level;

[0034] The target directional detection frame output module is used to use the detection head to perform target detection according to the multi-scale pyramid feature map to obtain the target directional detection frame to achieve directional detection of the ship target.

[0035] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0036] Acquire SAR measured images with ships as targets;

[0037] A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0038] A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0039] A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

[0040] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0041] Acquire SAR measured images with ships as targets;

[0042] A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0043] A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0044] A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

[0045] The above-mentioned SAR ship directional detection method, device, equipment and medium based on deep learning extract features from SAR measured images by using a multi-scale large convolution kernel feature extraction network to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes multiple multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer. Then, a feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes multiple feature fusion layers, and each feature fusion layer dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate a discriminative multi-scale pyramid feature map, and finally a detection head is used to output an accurate directional detection frame of the ship target in the SAR measured image according to the multi-scale pyramid feature map. Sampling This method can effectively improve the accuracy and efficiency of ship target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flowchart of a SAR ship directional detection method based on deep learning in one embodiment;

[0047] Figure 2 Schematic diagram of the structure of a multi-scale large convolution kernel feature extraction network in one embodiment;

[0048] Figure 3 Schematic diagram of the structure of a multi-scale large convolution kernel feature extraction layer in one embodiment;

[0049] Figure 4 is a schematic diagram of the structure of a feature fusion layer in one embodiment;

[0050] Figure 5 A schematic diagram of the conversion of the bounding box representation method and the loss measurement in one embodiment;

[0051] Figure 6 This is a schematic diagram showing the comparison of the detection visualization results of different training methods in the R-SSDD dataset in an experiment, where: Figure 6 (a) Figure 6 (b) and Figure 6 (c) is the detection visualization results of different training methods in the same scene. Figure 6 (d) Figure 6 (e) and Figure 6 (f) is the detection visualization result of using different training methods in another same scene. Figure 6 (c) and Figure 6 (f) is the detection visualization result using the method in this paper;

[0052] Figure 7 It is a structural block diagram of a SAR ship directional detection device based on deep learning in one embodiment;

[0053] Figure 8 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0055] Ship detection in SAR images in any orientation helps to extract target features more accurately and comprehensively in different scenarios. However, achieving optimal detection performance remains challenging due to multiple levels of imbalance, such as target scale imbalance, feature utilization imbalance, and regression loss imbalance. In order to address these challenges and thus achieve optimal performance for ship target detection, in this application, Figure 1 As shown, a SAR ship directional detection method based on deep learning is provided, which specifically includes the following steps:

[0056] Step S100, obtaining a SAR measured image with the ship as the target.

[0057] Step S110, using a multi-scale large convolution kernel feature extraction network to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer.

[0058] Step S120, using a feature fusion network to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes multiple feature fusion layers, and each feature fusion layer dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination.

[0059] Step S130, using a detection head to perform target detection according to the multi-scale pyramid feature map, and obtaining a target directional detection frame to achieve directional detection of the ship target.

[0060] In this method, a multi-scale dynamic feature fusion network for SAR ship detection in any direction is proposed. First, in order to alleviate the impact of scale imbalance, a new backbone framework MSLK-Net, i.e., a large-scale convolution kernel feature extraction network, is designed, which includes multiple multi-scale large-kernel convolution blocks (MSLK-Block). In this network, large-kernel convolution is combined with partitioned heterogeneous convolution to enhance the robust representation ability of the backbone network for multi-scale features, and can more effectively extract multi-scale ship target features and local and global information from complex scenes. Secondly, in order to enhance the perception ability of the model in complex scenes and solve the problem of unbalanced utilization of spatial features and channel features in the feature fusion process, an adaptive feature fusion module DFF, i.e., a feature fusion network, based on global context information is constructed in combination with the concept of dynamic mechanism. Subsequently, the DFF module is integrated with MSLK-Block to establish a new feature pyramid structure DFF-Net. This method can improve the detection efficiency while accurately predicting the directed detection box of ship targets in SAR images, thereby effectively improving the SAR ship target detection performance.

[0061] Specifically, after obtaining the SAR measured image in step S100, the SAR image is first subjected to multi-scale feature extraction by the multi-scale large convolution kernel feature extraction network MSLK-Net in step S110, and then the extracted multi-scale features are subjected to multi-scale feature fusion by the feature fusion network DFF-Net in step S120, and finally, in step S130, the detection head is used to perform directional target detection based on the multi-scale fusion features. Next, MSLK-Net and DFF-Net are introduced one by one.

[0062] When designing the core multi-scale large convolution kernel feature extraction network, it was observed that the receptive field size has a great impact on the detection results, especially for ship targets with multi-scale variability. Typical target detection models use a uniform convolution kernel size in the backbone, which has limitations: for small ships smaller than the receptive field, feature extraction may be insufficient, while for large ships larger than the receptive field, too much background information may obscure the target. This limitation in fixed kernel convolution affects the overall performance of the model. To address this issue, MSLK-Net adaptively adjusts the size of the large kernel convolution in each MSLK-Block according to the structural scale of the feature map in the backbone, such as Figure 2 Specifically, multiple large-depth kernels of different scales are used to capture the multi-scale features of ships. Instead of directly merging different large-depth kernels, MSLK-Net uses heterogeneous depth convolution to dynamically adjust the kernel size of each MSLK-Block.

[0063] In this embodiment, in the multi-scale large convolution kernel feature extraction network, the SAR measured image is first processed through the starting layer (Stem) to obtain an initial feature map, and then the initial feature map is input into the multi-scale large convolution kernel feature extraction layer for further feature extraction.

[0064] like Figure 2 As shown, in one embodiment, the multi-scale large convolution kernel feature extraction network includes a 5-layer network structure, first passing through three Conv modules That is, the starting layer, generating feature maps .Then The subsequent network is passed to four sequentially connected multi-scale large convolution kernel feature extraction layers to obtain , , and .

[0065] Since, in this embodiment, the multi-scale large convolution kernel feature extraction network includes a 5-layer network structure, including four layers of multi-scale large convolution kernel feature extraction layers, the kernel size of each MSLK-Block is set to 3, 5, 7 and 9 respectively. At the shallower level, in order to extract fine-grained features, 3×3 convolution is used. As the layer depth increases, the kernel size gradually grows, and eventually a larger convolution kernel is produced at a higher feature extraction level. This large kernel effectively captures a wider range of contextual information, which is conducive to the detection of larger objects, and the expanded receptive field enhances the model's spatial understanding around the target object, which is crucial for accurate detection and positioning. The deeper and more adaptable receptive field provided by the heterogeneous large kernel convolution enables the backbone to capture more subtle and information-rich features. Experimental results show that this structure not only improves the efficiency of convolution and reduces the computational cost, but also expands the receptive field of high-resolution and low-resolution feature maps, effectively extracting richer multi-scale features.

[0066] In the field of general target detection, some schemes have implemented large kernel convolution to enhance the model's capabilities in multi-scale ship target feature extraction. For example, RTMDet introduced a 5×5 large kernel convolution into the network structure to enhance the model's extraction function for each core module in its proposed backbone architecture. Although the proposed large kernel convolution method has certain advantages in computational efficiency and feature extraction, its fixed kernel size limits the ability to capture change information from the target. Considering the significant scale changes of ship targets, the information contained in SAR images is inherently limited and easily interfered by background noise. In order to solve this problem, in this method, large kernel convolution is combined with a multi-scale partitioning mechanism to construct a new MSLK-Layer, i.e., a multi-scale large convolution kernel feature extraction layer, as shown in Fig. Figure 3 shown.

[0067] In this embodiment, in the MSLK-Layer, a partitioned convolution method is introduced. For the features of the lower branch, an inverse bottleneck (IBN) layer is used. The IBN layer expands the number of input channels using a smaller convolution kernel (usually a 1×1 convolution) and then reduces it to the original number using another 1×1 convolution. This greatly reduces the number of parameters, thereby reducing the computational cost. In addition, in the inverse bottleneck layer, by introducing a large-scale convolution layer with a kernel size of between the expansion layer and the compression layer, the nonlinear mapping capability is enhanced to capture the changing multi-scale features.

[0068] Specifically, a partitioned convolution method is used in each multi-scale large convolution kernel feature extraction layer to send the input feature map to the upper branch and the lower branch respectively. In the lower branch, the input feature map is processed by convolution and then passes through a large convolution kernel feature extraction block for feature extraction. After the extracted feature map is fused with the feature map that passes through the upper branch, the output feature map is obtained through the attention channel and the convolution layer. The convolution kernel size of the large convolution kernel feature extraction block is determined by the number of layers in the multi-scale large convolution kernel feature extraction network; wherein, in the large convolution kernel feature extraction block, the feature map after convolution processing is feature grouped, and then the grouped feature map is processed using the reverse bottleneck layer to obtain the output feature map of the lower branch.

[0069] Specifically, in each multi-scale large convolution kernel feature extraction layer, refer to Figure 3 , first use 1×1 convolution to process the input features , which aims to change the number of channels in the extracted features to Then, the adjusted features The kernel size is The MSLK-Block is used for further feature extraction, where Determined by the number of layers of the backbone network, The processing process is expressed as:

[0070] (1)

[0071] In formula (1), Refers to the standard 1×1 convolution.

[0072] Furthermore, in the large convolution kernel feature extraction block: first, the grouped feature maps Perform group processing and Features are divided into For different groups, use the following formula:

[0073] (2)

[0074] In formula (2), Represents block features, Usually equal to 1, 2 and 3, Represents a grouping operation.

[0075] In order to enhance the relationship between different groups of features, an iterative processing method is used to group the features of each group. Processing is performed and the spatial relationships between these features are effectively extracted.

[0076] Specifically, for the first set of feature maps Remain unchanged.

[0077] Specifically, the first set of feature maps And the second set of feature maps Fusion is performed, and then the reverse bottleneck layer is used for feature extraction to obtain the first fusion feature First, we add and Fusion to obtain , then through the block That is, the reverse bottleneck layer performs feature extraction to obtain .

[0078] Further, the reverse bottleneck layer It consists of three parts: the first module is the dilated convolution, which adjusts the number of feature channels to obtain .Then, Through the large kernel depth separable convolution, first perform depth convolution, then perform point-by-point convolution to obtain With output channel Finally, the channel is adjusted to ,get The specific implementation process is expressed as:

[0079] (3)

[0080] (4)

[0081] (5)

[0082] (6)

[0083] In formula (3) to formula (6), is the expansion factor, and the number of output channels of the expansion layer is , It represents the element-wise addition of features.

[0084] Next, the first fusion feature map And the third set of feature maps Fusion, and then use the reverse bottleneck layer Perform feature extraction to obtain the second fusion feature Similarly, the specific implementation process is expressed as:

[0085] (7)

[0086] (8)

[0087] (9)

[0088] (10)

[0089] Afterwards, the first set of feature maps , the first fusion feature map And the second fusion feature map The input features of the large convolution kernel feature extraction block are obtained by combining them along the channel axis and feeding them into the channel attention mechanism for dynamic adjustment at the channel level.

[0090] Specifically, the process of obtaining the input feature is:

[0091] (11)

[0092] (12)

[0093] (13)

[0094] (14)

[0095] (15)

[0096] In formula (11) to formula (15), Represents the concatenation of features along the channel dimension. represents the adaptive average pooling module, Represented as a fully connected layer, represents the classic HardSigmoid function, represents the linear activation function, Represents element-wise multiplication.

[0097] at last, Output features are generated through standard 1×1 convolution , whose dimensions are .

[0098] In this embodiment, the above-mentioned multi-scale large convolution kernel feature extraction layer has two significant advantages. First, by partitioning the features, it captures the inter-block dependencies in complex scenes, effectively retains multi-level feature information and achieves efficient multi-scale feature fusion. Second, the IBN-based feature extraction method enhances the extraction of key features while maintaining the computational efficiency of large kernel convolution.

[0099] At present, most SAR ship detection models use static, horizontal fusion in feature fusion, and all input images share the same fusion weights. However, it is observed that this static method cannot take into account the scale changes and local target changes of SAR images, thereby limiting the adaptability of feature fusion. In addition, these methods often ignore the correlation between scales, which hinders the extraction of discriminative fusion features in complex scenes. In addition, over-emphasizing channel correlation ignores the role of spatial local features in channel-based modeling. In order to solve these problems, in this embodiment, a DFF-Block, i.e., a feature fusion layer, is proposed to balance the use of spatial and channel features in the feature fusion process. The feature fusion layer uses global information complementarity to enhance the integration of multi-scale backbone features and generates a multi-scale feature map rich in semantic information and detailed information. A feature fusion network is constructed based on multiple feature fusion layers, and the multi-scale features are fused using the network.

[0100] In this embodiment, in order to balance computational efficiency and detection performance, the first three layers of feature maps of different scales obtained by the multi-scale large convolution kernel feature extraction network are input into the feature fusion network for feature fusion enhancement. Processed by MSLK-Net to collect features at different levels In order to balance computational efficiency and detection performance, the first three layers of features are selected Perform feature fusion.

[0101] Specifically, in When performing feature fusion, the intermediate layer features are also obtained through downsampling and channel concatenation , the intermediate layer features Input into the feature fusion network for feature fusion enhancement.

[0102] like Figure 4 As shown in the figure, in the feature fusion layer, the feature fusion method used utilizes the global spatial scale correlation between high-level and low-level feature maps in the backbone network to dynamically optimize channel features to generate more discriminative multi-scale pyramid features. It mainly includes two parts: spatial feature adaptation and channel feature optimization.

[0103] Specifically, at each feature fusion layer, the spatial and channel scales of the high-level scale feature map and the low-level scale feature map are adjusted to obtain the high-level scale input feature map and the low-level scale input feature map. Assume that the input feature is , and obtain multi-scale features from it through the backbone network. Taking the kth and jth layers as examples, and Represent the high-level and low-level scale features extracted by the backbone network respectively. Due to the differences in spatial and channel scales between features, it is necessary to adjust before feature fusion and The spatial and channel scales of the input features and .

[0104] Furthermore, in the global channel feature optimization stage, the high-level scale is input into the feature map And the low-level scale input feature map After fusion, based on the fusion feature Get the global channel attention weight. The fusion process is expressed as:

[0105] (16)

[0106] (17)

[0107] (18)

[0108] In traditional static feature fusion methods, fusion features are usually obtained by concatenation. In this embodiment, As the basic fusion feature, on this basis, the importance of the fusion feature is further adjusted at the spatial and channel levels.

[0109] Specifically, in the global channel feature optimization stage, the fusion feature Perform global average pooling (AvgPool) to obtain the global spatial attention vector Then for this vector conduct Convolution operation, adjust the weights between different channels, and obtain the convolution vector Finally, for this vector Perform Sigmoid function processing to normalize the value to the range of [0, 1] to obtain the global channel attention weight , the process is expressed as:

[0110] (19)

[0111] (20)

[0112] (twenty one)

[0113] In formula (19) to formula (21), express In the channel and spatial location The eigenvalue at express The feature value after global average pooling, express Convolution operation.

[0114] Furthermore, in the global spatial feature optimization stage, convolution is used to process the high-level scale input feature map and the low-level scale input feature map to adjust the corresponding weights, and the two weights are added element-by-element to obtain the global spatial attention weight.

[0115] Specifically, we first use 1×1 convolution to process the two input feature branches to adjust the corresponding weights. and The two weights are then added element-wise to produce the global spatial attention weight .

[0116] Finally, according to the global channel attention weight , global spatial attention weight And fusion features Get the corresponding scale pyramid feature map output by the feature fusion layer, where the fusion feature It is the feature obtained by fusing the high-level scale input feature map and the low-level scale input feature map.

[0117] Specifically, the fusion features and global channel attention weights After element-by-element multiplication, the convolution operation is performed to obtain the globally weighted fusion feature. , and then use the global spatial attention weight With fusion features Perform element-by-element multiplication to obtain the corresponding scale pyramid feature map output by the feature fusion layer .

[0118] In the above feature fusion network, the feature map , and Through DFF-Block dynamic fusion, the enhanced feature map is obtained , and Without significantly increasing the complexity of the model, balanced optimization of global channel and spatial features across multi-scale pyramid features is achieved. Subsequent experimental results show that this enhancement can improve the overall detection effect of the model, thereby improving the accuracy of identifying ship targets in various scenarios.

[0119] Finally, in the detection head in step S130, according to the feature fusion network optimized Make result predictions and output the directed detection box of the ship target.

[0120] In this embodiment, a ship target directional detection network can be constructed based on the above-mentioned multi-scale large convolution kernel feature extraction network, feature fusion network and detection head, that is, the multi-scale dynamic feature fusion network for SAR ship detection in any direction mentioned above, and the SAR measured image can be input into the ship target directional detection network for target directional detection.

[0121] In this embodiment, when constructing a training data set for a directed ship target detection network: a data set with a rotation bounding box annotation is obtained, the data set includes a plurality of SAR images with ships as targets, the data set is expanded to obtain an enhanced data set, and the enhanced data set is used as a training data set.

[0122] Specifically, assume that the original training data set is recorded as ,in Defined as the number of images used during training. Each image There is a corresponding OBB (with direction) annotation, represented as By obtaining an enhanced dataset , For data enhancement operations, Represents the number of images in the augmented dataset.

[0123] In this embodiment, when training the ship target directed detection network, the boundary regression loss is used. and classification loss Training is performed where the boundary regression loss is formulated by optimizing the loss in conjunction with a distance constraint between Gaussian distributions.

[0124] like Figure 5 As shown, the conversion process of the bounding box representation method and the corresponding loss calculation are described.

[0125] Specifically, when calculating the boundary regression loss, first, represent OBB Convert to a specific 2-D GPD distribution .here represents the center coordinates of the true value of the OBB, and Indicates the corresponding length and width, Indicates the rotation angle of the OBB.

[0126] By the mean vector and the covariance matrix express. corresponds to the center of the OBB and is calculated using the following formula:

[0127] (twenty two)

[0128] Covariance matrix encapsulates the spread and orientation of the distribution. For bounding boxes without rotation ( ), the covariance matrix is ​​initially expressed as:

[0129] (twenty three)

[0130] In formula (23), represents the unrotated covariance, where and denote the standard deviation along the major and minor axes, respectively. To explain the rotation angle , apply the rotation matrix to transform the covariance matrix:

[0131] (twenty four)

[0132] Then we get the rotated covariance matrix , expressed as:

[0133] (25)

[0134] Substituting in the values, the covariance matrix is ​​given by:

[0135] (26)

[0136] Assumptions represents the OBB prediction parameters of network regression, represents the OBB label, then the corresponding GBB distributions are and The Bhattacharyya coefficient between two Gaussian distributions Distance from Bhattacharyya The calculation is as follows:

[0137] (27)

[0138] (28)

[0139] In formula (27) and formula (28), Represents the determinant function, which is a scalar value that can be calculated using a square matrix.

[0140] The Bhattacharyya distance takes into account the center position and shape of the distribution, thus more accurately measuring the similarity between two two-dimensional Gaussian distributions. It cannot be used directly for network regression and must be further converted into Hellinger distance :

[0141] (29)

[0142] Then, based on the ProbIOU loss calculation standard, the distribution metric regression loss is obtained as follows:

[0143] (30)

[0144] (31)

[0145] here, and They represent the regression loss values ​​at different stages of training.

[0146] Specifically, QFL is used to calculate the classification loss :

[0147] (32)

[0148] In formula (32), represents the number of positive instances, Represents the IoU value between the predicted bounding box and the annotation box, represents the probability value predicted by the corresponding model, It is the attenuation factor for adjusting the difficulty of samples. Preferably, it is set to 2.0 to control the weights of easy samples and difficult samples.

[0149] Then the overall network training loss calculation formula is expressed as:

[0150] (33)

[0151] In formula (33), represents the total loss value, and It is a training adjustment factor that dynamically adjusts the regression loss component according to the number of training cycles.

[0152] In this paper, the effectiveness of the proposed method is also demonstrated through experimental data. First, the three datasets used in the experiment, the evaluation criteria and the corresponding experimental parameter settings are introduced. Then, detailed ablation experimental comparison results are given, as well as corresponding visualization, quantitative experimental results and PR curve analysis.

[0153] In the experiment, the SSDD dataset is used. This dataset is the first dataset created specifically for detecting ship targets in SAR images, with HBB and OBB annotation formats. The dataset contains images from three radar satellites: RadarSat-2, TerraSAR-X, and Sentinel-1, with resolutions ranging from 1 to 15 meters. Ship targets are distributed in offshore and nearshore scenes. The original dataset includes 1160 images and 2456 ship instances, with image sizes ranging from 190 pixels to 688 pixels. In order to standardize the input image size, the original dataset was processed and all images were resized to 512 pixels × 512 pixels, resulting in a total of 1247 images (872 images for training and 375 for testing). In addition, the annotations of this dataset were modified based on expert experience, resulting in the R-SSDD used in our experiments. R-SSDD contains a total of 2853 ship instances (1975 for training and 878 for testing).

[0154] In the experiment, precision, recall, average precision, and F1-Score are used to evaluate the detection accuracy performance of the proposed method. In addition, the number of parameters (Params), computational complexity (FLOPS), and computational efficiency (FPS) are used to comprehensively analyze the complexity and computational efficiency of the above network.

[0155] All comparative experiments (including anchor-free, one-stage, and two-stage detection algorithms) were conducted using the basic settings of MMRotate. The backbone network was initialized with pre-trained weights from ImageNet to accelerate model convergence. The models were trained using AdamW or SGD optimizers. The initial setting of the model learning rate of AdamW was 0.0001, and the regularization term was 0.0005. For SGD, the initial learning rate was 0.001 or 0.005, the momentum parameter was 0.9, and the regularization coefficient was 0.0001. The MultiStepLR method was used to automatically adjust the learning rate in the experiment. The input size of the detection algorithm was adjusted according to the scale of the input dataset.

[0156] Ablation experiment and analysis:

[0157] Effect of MSLK-Net: RTMDet-OBB is established as the basic baseline model for ablation studies. Based on the baseline algorithm, the module scheme proposed in this paper is gradually adopted, and a comparative analysis is performed to emphasize the contribution of each component.

[0158] Table 1 shows the quantitative experimental results of the proposed backbone model and the baseline model in different scenarios in the R-SSDD dataset. From the comparison results, MSLK-Net significantly improves the detection performance in nearshore and offshore scenarios. For the nearshore scenario, feature extraction using MSLK-Net can achieve 85.21% detection accuracy and 84.95% recall, and mAP and F1-Score reach 81.47% and 85.08%, respectively. Compared with the baseline model, these indicators are improved by 0.89%, 1.34%, 1.25%, and 1.12%, respectively. For the offshore scenario, compared with the baseline model, the proposed feature extraction method improves the four indicators by 0.72%, 0.68%, 0.81%, and 0.70%, reaching 95.47%, 96.62%, 91.03%, and 96.04%, respectively. The results in Table 1 further verify that the proposed MSLK-Net improves the model's ability to extract features from ships of different scales, thereby improving the target detection performance.

[0159] Table 1 R-SSDD effect of MSLK-NET in different scenarios (%)

[0160]

[0161] Effect of DFF-Net: In addition to optimizing the feature extraction module, the proposed algorithm also improves the feature fusion process. By adding a dynamic fusion mechanism, the DFF-Net in this method dynamically integrates the features extracted from the backbone models at different levels, effectively enhancing the semantic association between global and local scene information. In order to evaluate the effectiveness of the feature fusion method, the fused features were visually analyzed using Eigen-CAM. Eigen-CAM generates multi-scale category activations (CAMs) from feature maps at different levels, which can be directly applied to pre-trained models and is very suitable for verifying the feasibility of the DFF-Net structure. It can be observed that DFF-Net achieves better spatial scale feature fusion compared with the baseline model. This is particularly evident in the nearshore scene, where the fused features focus more on the target area and less on land interference. This shows that the proposed DFF-Net model enhances the focus on the target between global and local semantics and improves the detection performance in complex scenes. The experimental results confirm that the proposed feature fusion method effectively suppresses land interference in complex scenes, enhances scene-level semantic information, and improves model detection performance.

[0162] Impact of GPD loss: In order to better evaluate the effectiveness of different prediction methods for angle estimation, the mean angle error (MAE) is used in the experiment to verify the impact of the proposed metric on angle prediction. MAE is calculated as follows:

[0163] (34)

[0164] In formula (34), represents the angle corresponding to the i-th predicted bounding box, Indicates

[0165] The angles corresponding to the i ground truth bounding boxes, and Indicates the total number of samples tested.

[0166] The calculation of the angle error must take into account periodicity. For example, in the case of the long edge definition, 180° and 0° should be considered as close angles. Therefore, the minimum angle difference is usually used to calculate the angle error:

[0167] (35)

[0168] Table 2 shows the angle prediction errors of the two prediction methods across different scenarios on different datasets. In the baseline model, an angle prediction vector is added and the rotation IoU loss is directly calculated for regression as the comparison method in this paper. As shown in Table 2, for ships in near-shore scenarios, the method based on rotation IoU loss obtains MAE values ​​of 0.1222, 0.1098, and 0.0651 on the three datasets, respectively. In contrast, our proposed method uses a Gaussian-based metric combined with ProbIOU loss, and the MAE values ​​obtained on these datasets are only 0.0870, 0.1023, and 0.0542. The MAE is reduced by 0.0352, 0.0075, and 0.0109 (angles converted to radians). For ocean-going ships, the network in this paper achieves MAE values ​​of 0.0846, 0.0530, and 0.0438 on the three datasets, which are significantly lower than the comparison methods of 0.0893, 0.0548, and 0.0522. These quantified angle prediction errors clearly illustrate the effectiveness of the Gaussian-based angle prediction and loss calculation method, which reduces the computational complexity while improving the prediction accuracy. In addition, the visualization detection results of the two methods in the nearshore scene of the R-SSDD dataset are provided, such as Figure 6 As shown in the figure, the ellipse represents a false alarm, and the third column is the detection performance of the algorithm proposed in this paper. It can be seen from the figure that the head prediction method proposed in this paper not only provides a more accurate oriented bounding box, which can better fit the target, but also shows significant advantages in terms of overall detection precision recall.

[0169] Table 2 Comparison of angle prediction errors of different loss training methods (radians)

[0170]

[0171] In the above deep learning-based SAR ship directional detection method, in order to reduce the impact of scale imbalance on feature extraction, a multi-scale large kernel convolution block is introduced into the backbone structure, and a new backbone framework MSLK-Net is proposed. MSLK-Net can more effectively extract multi-scale ship target features and local and global information from complex scenes. In order to enhance the perception ability of the model in complex scenes and solve the problem of unbalanced utilization of spatial features and channel features in the feature fusion process, an adaptive feature fusion module DFF based on global context information is constructed in combination with the concept of dynamic mechanism. Subsequently, the DFF module is integrated with MSLK-Block to establish a new feature pyramid structure DFF-Net. Inspired by the elliptical characteristics of ship targets and directional target detection, a ship target feature fusion module DFF based on dynamic mechanism is proposed. The operation of converting the OBB representation to a Gaussian distribution is performed in the prediction head. For loss calculation, an improved probability IoU is introduced to calculate the target regression loss. By measuring the GPD distance, this method guides the optimization of the network model, thereby reducing the imbalance between regression and classification tasks in the model learning process.

[0172] In this method, a SAR ship detection method for multiple scenarios in any direction, MSDFF-Net, is proposed. This method can effectively suppress the impact of the multi-level imbalance problem in SAR ship detection, and shows strong performance in detection accuracy, model parameters, and computational efficiency. Two feature balancing modules, MSLK-Block and DFF-Block, are constructed. The former achieves a balance in target-level scale feature extraction by integrating the advantages of large kernels and block heterogeneous convolution. The latter achieves a balance in spatial and channel feature fusion utilization under dynamic feature adaptive adjustment. In order to solve the imbalance problem between regression and classification losses, the GBB fitting method is introduced. The loss constraint is achieved by calculating the Bhattacharyya distance between the target Gaussian probability distributions, which provides better decision support for accurate directional prediction of ships. The proposed method is extensively experimentally evaluated on three SAR ship target detection datasets with different resolutions, and a standardized experimental benchmark is established. The effectiveness of the proposed method is thoroughly verified by various quantitative indicators and feature visualization results.

[0173] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0174] In one embodiment, Figure 7 As shown, a SAR ship directional detection device based on deep learning is provided, comprising: a SAR measured image acquisition module 200, a multi-scale feature extraction module 210, a multi-scale fusion feature extraction module 220 and a target directional detection frame output module 230, wherein:

[0175] The SAR measured image acquisition module 200 is used to acquire a SAR measured image with the ship as the target;

[0176] A multi-scale feature extraction module 210 is used to extract features from the SAR measured image using a multi-scale large convolution kernel feature extraction network to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0177] A multi-scale fusion feature extraction module 220 is used to use a feature fusion network to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, each of which dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0178] The target directional detection frame output module 230 is used to use a detection head to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of the ship target.

[0179] For the specific definition of the SAR ship directional detection device based on deep learning, please refer to the definition of the SAR ship directional detection method based on deep learning above, which will not be repeated here. Each module in the above-mentioned SAR ship directional detection device based on deep learning can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0180] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a SAR ship directional detection method based on deep learning is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0181] Those skilled in the art will understand that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0182] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the following steps are implemented:

[0183] Acquire SAR measured images with ships as targets;

[0184] A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0185] A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0186] A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

[0187] In one embodiment, a computer readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the following steps are implemented:

[0188] Acquire SAR measured images with ships as targets;

[0189] A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer;

[0190] A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination;

[0191] A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

[0192] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0193] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A SAR ship directional detection method based on deep learning, characterized in that: The method comprises: Acquire SAR measured images with ships as targets; A multi-scale large convolution kernel feature extraction network is used to extract features from the SAR measured image to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and a heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer. The SAR measured image is first processed through a starting layer to obtain an initial feature map, and then the initial feature map is input into the multi-scale large convolution kernel feature extraction layer. A partitioned convolution method is used in each of the multi-scale large convolution kernel feature extraction layers to respectively input the input feature map into The upper branch and the lower branch, in the lower branch, the input feature map is subjected to convolution processing and then passes through a large convolution kernel feature extraction block for feature extraction, the extracted feature map is fused with the feature map that passes through the upper branch, and then the output feature map is obtained through the attention channel and the convolution layer, the convolution kernel size of the large convolution kernel feature extraction block is determined by the number of layers in the multi-scale large convolution kernel feature extraction network, wherein, in the large convolution kernel feature extraction block, the feature map after convolution processing is feature grouped, and then the grouped feature map is processed by the reverse bottleneck layer to obtain the output feature map of the lower branch; A feature fusion network is used to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, and each of the feature fusion layers dynamically optimizes channel features by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level to generate the multi-scale pyramid feature map with discrimination; A detection head is used to perform target detection according to the multi-scale pyramid feature map to obtain a target directional detection frame to achieve directional detection of ship targets.

2. The SAR ship directional detection method according to claim 1, characterized in that: In the large convolution kernel feature extraction block: The grouped feature maps are processed using an iterative processing method, and the first group of feature maps remain unchanged; The first set of feature maps and the second set of feature maps are fused, and then a reverse bottleneck layer is used to extract features to obtain a first fused feature map; The first fused feature map and the third set of feature maps are fused, and then a reverse bottleneck layer is used to perform feature extraction to obtain a second fused feature map; The first group of feature maps, the first fused feature map and the second fused feature map are combined along the channel axis and input into the channel attention mechanism for dynamic adjustment at the channel level to obtain input features of the large convolution kernel feature extraction block.

3. The SAR ship directional detection method according to claim 1 or 2, characterized in that: The feature maps of different scales in the first three layers obtained by the multi-scale large convolution kernel feature extraction network are input into the feature fusion network for feature fusion enhancement.

4. The SAR ship directional detection method according to claim 3, characterized in that: In each of the feature fusion layers: Adjusting the spatial and channel scales of the high-level scale feature map and the low-level scale feature map to obtain a high-level scale input feature map and a low-level scale input feature map; In the global channel feature optimization stage, the high-level scale input feature map and the low-level scale input feature map are fused, and a global channel attention weight is obtained based on the fused feature; In the global spatial feature optimization stage, convolution is used to process the high-level scale input feature map and the low-level scale input feature map to adjust the corresponding weights, and the two weights are added element by element to obtain the global spatial attention weight; A corresponding scale pyramid feature map output by the feature fusion layer is obtained according to the global channel attention weight, the global spatial attention weight and the fusion feature, wherein the fusion feature is a feature obtained by fusing the high-level scale input feature map and the low-level scale input feature map.

5. The SAR ship directional detection method according to claim 4, characterized in that: A ship target directional detection network is constructed according to the multi-scale large convolution kernel feature extraction network, the feature fusion network and the detection head, and the SAR measured image is input into the ship target directional detection network for target directional detection. When constructing a training data set for the ship target directional detection network: Acquire a data set with rotation bounding box annotations, wherein the data set includes a plurality of SAR images with a ship as a target; The data set is expanded to obtain an enhanced data set, and the enhanced data set is used as the training data set.

6. The SAR ship directional detection method according to claim 5, characterized in that: When training the ship target directed detection network, boundary regression loss and classification loss are used for training, wherein the boundary regression loss is represented by optimizing the loss by combining the distance constraint between Gaussian distributions.

7. A SAR ship directional detection device based on deep learning, characterized in that: The device comprises: SAR measured image acquisition module, used to acquire SAR measured images with ships as targets; The multi-scale feature extraction module is used to extract features from the SAR measured image by using a multi-scale large convolution kernel feature extraction network to obtain a multi-scale feature map, wherein the multi-scale large convolution kernel feature extraction network includes a plurality of multi-scale large convolution kernel feature extraction layers connected in sequence, and heterogeneous deep convolution is used to dynamically adjust the kernel size of each multi-scale large convolution kernel feature extraction layer, the SAR measured image is first processed through the starting layer to obtain an initial feature map, and then the initial feature map is input into the multi-scale large convolution kernel feature extraction layer, and a partitioned convolution method is used in each of the multi-scale large convolution kernel feature extraction layers to extract the input feature map. The feature maps are sent to the upper branch and the lower branch respectively. In the lower branch, the input feature map is subjected to convolution processing and then passes through a large convolution kernel feature extraction block for feature extraction. After the extracted feature map is fused with the feature map that passes through the upper branch, an output feature map is obtained through an attention channel and a convolution layer. The convolution kernel size of the large convolution kernel feature extraction block is determined by the number of layers in the multi-scale large convolution kernel feature extraction network. In the large convolution kernel feature extraction block, the feature map after convolution processing is feature grouped, and then the grouped feature map is processed using a reverse bottleneck layer to obtain an output feature map of the lower branch. A multi-scale fusion feature extraction module, configured to use a feature fusion network to perform dynamic feature fusion according to the multi-scale feature map to obtain a multi-scale pyramid feature map, wherein the feature fusion network includes a plurality of feature fusion layers, each of which dynamically optimizes channel features to generate the multi-scale pyramid feature map with discriminability by utilizing the global spatial scale correlation between the high-level scale feature map and the low-level scale feature map of the corresponding level; The target directional detection frame output module is used to use the detection head to perform target detection according to the multi-scale pyramid feature map to obtain the target directional detection frame to achieve directional detection of the ship target.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images