Target segmentation method, device and server for side-scan sonar wreck image

Abstract

The application provides a side-scan sonar shipwreck image target segmentation method and device and a server, relates to the technical field of sonar image segmentation, and comprises the following steps: obtaining a shipwreck image of a side-scan sonar, performing multi-scale feature extraction on the shipwreck image by using a hybrid hollow convolution model to obtain a target feature region positioning map, performing attention weight analysis on the shipwreck image by using an attention segmentation model to obtain a multi-scale feature attention weighted response map, training an image segmentation model according to the target feature region positioning map and the multi-scale feature attention weighted response map to obtain a target image segmentation model, and performing segmentation processing on the shipwreck image by using the target image segmentation model to obtain a target segmentation image. The application can significantly reduce image segmentation errors and improve the target segmentation effect of the side-scan sonar shipwreck image.

Description

Technical Field

[0001] This invention relates to the field of sonar image segmentation technology, and in particular to a target segmentation method, apparatus and server for side-scan sonar images of shipwrecks. Background Technology

[0002] Traditional side-scan sonar images are interpreted manually, which is inefficient, time-consuming, and highly dependent on experience. Currently, related technologies have proposed using models such as Faster R-CNN, YOLOv5a, and lightweight DETR-YOLO to detect shipwreck targets. However, this approach has a high false alarm rate and missed alarm rate when detecting multiple closely arranged and overlapping shipwreck targets, which cannot meet the needs of practical engineering tasks. If Deeplab semantic segmentation models are used to segment images, the complex model structure and excessive number of parameters can easily lead to overfitting, failing to achieve the desired segmentation effect. Summary of the Invention

[0003] In view of this, the purpose of the present invention is to provide a target segmentation method, apparatus and server for side-scan sonar shipwreck images, which can significantly reduce image segmentation errors and improve the target segmentation effect of side-scan sonar shipwreck images.

[0004] In a first aspect, embodiments of the present invention provide a target segmentation method for side-scan sonar images of shipwrecks. The method is applied to an image segmentation processing system, which is configured with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolution model and an attention segmentation model. The method includes: acquiring a side-scan sonar image of a shipwreck, wherein the shipwreck image includes multiple overlapping shipwrecks; using the hybrid dilated convolution model to extract multi-scale features from the shipwreck image to obtain a target feature region localization map; using the attention segmentation model to perform attention weight analysis on the shipwreck image to obtain a multi-scale feature attention-weighted response map; training the image segmentation model based on the target feature region localization map and the multi-scale feature attention-weighted response map to obtain a target image segmentation model; and using the target image segmentation model to segment the shipwreck image to obtain a target segmented image.

[0005] In one implementation, the step of using a hybrid dilated convolution model to extract multi-scale features from a shipwreck image to obtain a target feature region localization map includes: acquiring target recognition localization maps corresponding to various convolution dilation rates within a preset dilation rate threshold for the shipwreck image, and establishing a target recognition localization map set; and obtaining a target feature region localization map by fusing the various target recognition localization maps in the target recognition localization map set.

[0006] In one implementation, the step of fusing the target recognition and localization maps in the target recognition and localization map set to obtain a target feature region localization map includes: sorting the target recognition and localization maps in the target recognition and localization map set according to the order of convolution dilation rate from smallest to largest; and fusing the previous level target recognition and localization map with the next level target recognition and localization map step by step to obtain a target feature region localization map.

[0007] In one implementation, after the step of establishing the target recognition and localization map set, the method includes: using a preset noise-resistant fusion model to correct the target feature region localization map and remove non-key regions in the target recognition and localization map, wherein the non-key regions are regions that are generated in the target recognition and localization map corresponding to the convolution dilation rate during the expansion process and are unrelated to the shipwreck image.

[0008] In one implementation, the attention segmentation model includes: a weight analysis unit, which performs attention weight analysis on the shipwreck image using the attention segmentation model to obtain a multi-scale feature attention-weighted response map, including: acquiring a spatial multi-scale feature map of the shipwreck image; using the weight analysis unit to extract the attention vectors of the channels at each scale in the spatial multi-scale feature map to obtain attention weight information for each scale channel; and using the attention weight information to perform attention weighting processing on the spatial multi-scale feature map to obtain a multi-scale feature attention-weighted response map.

[0009] In one implementation, the attention segmentation model includes a segmentation connection unit and a step of obtaining a spatial multi-scale feature map of a shipwreck image, including: using the segmentation connection unit to segment the shipwreck image into channels to obtain a spatial multi-scale feature map.

[0010] In one implementation, the step of performing attention weighting processing on spatial multi-scale feature maps using attention weight information to obtain a multi-scale feature attention-weighted response map includes: correcting the attention weight information of each scale channel using a preset calibration calculation model to obtain target attention weight information; multiplying the target attention weight information with the spatial multi-scale feature map of the corresponding scale to obtain an updated target feature map; and stitching together the target feature maps of each scale to obtain a multi-scale feature attention-weighted response map.

[0011] Secondly, embodiments of the present invention also provide a target segmentation device for side-scan sonar shipwreck images. The device is applied to an image segmentation processing system, which is configured with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolution model and an attention segmentation model. The device includes: an image acquisition module for acquiring side-scan sonar images of shipwrecks, wherein the shipwreck images include multiple overlapping shipwrecks; a feature region localization module for extracting multi-scale features from the shipwreck images using the hybrid dilated convolution model to obtain a target feature region localization map; an attention concentration module for performing attention weight analysis on the shipwreck images using the attention segmentation model to obtain a multi-scale feature attention-weighted response map; a segmentation model training module for training the image segmentation model based on the target feature region localization map and the multi-scale feature attention-weighted response map to obtain a target image segmentation model; and an image segmentation module for segmenting the shipwreck images using the target image segmentation model to obtain a target segmented image.

[0012] Thirdly, embodiments of the present invention also provide a server, including a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement any of the methods provided in the first aspect.

[0013] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement any of the methods provided in the first aspect.

[0014] The embodiments of the present invention bring the following beneficial effects:

[0015] This invention provides a target segmentation method, apparatus, and server for side-scan sonar shipwreck images. After acquiring the shipwreck image from the side-scan sonar, the method uses a hybrid dilated convolution model to extract multi-scale features from the shipwreck image, obtaining a target feature region localization map. Then, it uses an attention segmentation model to perform attention weight analysis on the shipwreck image, obtaining a multi-scale feature attention-weighted response map. Based on the target feature region localization map and the multi-scale feature attention-weighted response map, the image segmentation model is trained to obtain a target image segmentation model. This target image segmentation model is then used to segment the shipwreck image, resulting in a segmented target image. This invention can significantly reduce image segmentation errors and improve the target segmentation effect of side-scan sonar shipwreck images.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a target segmentation method for side-scan sonar images of a shipwreck provided in an embodiment of the present invention;

[0020] Figure 2 A flowchart illustrating the process of determining a target feature region location map, provided by an embodiment of the present invention;

[0021] Figure 3 A flowchart illustrating an optimization method for an optimized training module provided in an embodiment of the present invention;

[0022] Figure 4 A schematic diagram of the structure of a target segmentation device for side-scan sonar images of a shipwreck provided in an embodiment of the present invention;

[0023] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Traditional side-scan sonar images are interpreted manually, which is inefficient, time-consuming, and highly dependent on experience. Current technologies suggest that models such as Faster R-CNN, YOLOv5a, and the lightweight DETR-YOLO can be used to detect shipwreck targets. If Faster R-CNN is used... The R-CNN model can automatically detect shipwreck targets using side-scan sonar, but it suffers from complex structure and low training and detection efficiency. While the YOLOv3 model improves training and detection efficiency, it still suffers from high false alarm rates for small targets and insufficient real-time detection speed. The YOLOv5a model improves accuracy and efficiency for small-scale targets, but its performance in complex sea conditions needs improvement. The lightweight DETR-YOLO model enhances overall scene understanding and detection capabilities in complex marine environments while meeting lightweight engineering deployment requirements. Although these algorithms have achieved good results in detecting shipwreck targets using side-scan sonar, they suffer from high false alarm and false alarm rates in detecting densely packed, overlapping, and complex multi-shipwreck targets, failing to meet the needs of practical engineering tasks.

[0026] In recent years, the emergence of deep convolutional neural networks has greatly promoted the development of semantic segmentation. It has developed rapidly in fields such as intelligent security, autonomous driving, satellite remote sensing, medical image processing, biometric recognition, VR and AR. Among them, the Deeplab semantic segmentation series models (Deeplabv1

[23] , Deeplabv2

[24] , Deeplabv3

[25] and Deeplabv3+

[26] ) are based on atrous spatial pyramid configuration. ASPP pooling expands the receptive field without changing the resolution and achieves feature fusion at different scales. The Deeplabv3+ model combines the advantages of the Encode-Decoder and ASPP structures, making it a semantic segmentation algorithm with excellent overall performance. However, due to its complex model structure and numerous parameters, the Deeplabv3+ model requires a large amount of data samples for training to achieve the desired segmentation effect. Therefore, its advantages are most obvious in semantic segmentation tasks with large data samples and high target complexity. Based on this, the target segmentation method for side-scan sonar shipwreck images provided in this invention is based on the image segmentation model Unet. It uses multi-scale hybrid dilated convolution and attention mechanisms to train Unet, thereby improving image segmentation performance and significantly reducing image segmentation errors, thus improving the target segmentation effect of side-scan sonar shipwreck images.

[0027] See Figure 1The diagram shows a flowchart of a target segmentation method for side-scan sonar images of a shipwreck. This method is applied to an image segmentation processing system, which is equipped with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolution model and an attention segmentation model. The method mainly includes the following steps S102 to S110:

[0028] Step S102: Obtain a side-scan sonar image of the shipwreck. The shipwreck image includes multiple overlapping shipwrecks. In one embodiment, the semantics of the shipwreck image are relatively simple and the structure is relatively fixed. The structure of the shipwreck target is relatively fixed under sonar imaging. Therefore, the importance of deep semantic information and low-level features is high. In addition, the amount of data in the shipwreck image is limited. There are relatively few shipwrecks on the seabed in reality. At the same time, the acquisition of sonar data of side-scan sonar shipwrecks is relatively difficult and costly.

[0029] Step S104: Using a hybrid dilated convolution model, multi-scale feature extraction is performed on the shipwreck image to obtain a target feature region localization map. The hybrid dilated convolution model is a multi-scale hybrid dilated convolution model (BHD Module), and the target feature region localization map is a complete localization map that highlights the target feature region by fusing localization regions generated by different dilation rates.

[0030] Step S106: Using the attention segmentation model, attention weight analysis is performed on the shipwreck image to obtain a multi-scale feature attention-weighted response map. The attention segmentation model includes a weight analysis unit and a segmentation connection unit. The multi-scale feature attention-weighted response map is used to specifically enable the algorithm model to focus on features that the user considers important. In one implementation, the attention segmentation model (PSA Module) may include three parts: a segmentation connection unit (SPC Module (Split and Concat Module)), a weight analysis unit (SEWeight Module), and Fscale, which is a lightweight and effective pyramid segmentation attention module.

[0031] Step S108: Based on the target feature region localization map and the multi-scale feature attention-weighted response map, the image segmentation model is trained to obtain the target image segmentation model. In one implementation, the original image segmentation model is trained using the analysis results of the side-scan sonar shipwreck image obtained by the hybrid dilated convolution model and attention segmentation model. This improves the receptive field while integrating the learning ability of deep semantics and shallow features, and establishes long-term dependencies between global and local information while processing multi-scale spatial features, thereby effectively improving the segmentation effect of the model.

[0032] Step S110: The shipwreck image is segmented using a target image segmentation model to obtain a target segmented image, wherein the target segmented image is a segmented image of multiple closely arranged and overlapping shipwrecks.

[0033] The target segmentation method for side-scan sonar shipwreck images provided in this embodiment of the invention can significantly reduce image segmentation errors and improve the target segmentation effect of side-scan sonar shipwreck images.

[0034] This invention also provides an implementation method for obtaining a target feature region localization map, see below. Figure 2 The flowchart shown is a method for determining the location map of a target feature region. For details, please refer to (1) to (2) below:

[0035] (1) Obtain the target recognition and localization map corresponding to each convolutional dilation rate within the preset dilation rate threshold of the shipwreck image, and establish a set of target recognition and localization maps. In one embodiment, a preset anti-noise fusion model is used to correct the target feature region localization map and remove non-key regions in the target recognition and localization map. The non-key regions are regions that are generated in the target recognition and localization map corresponding to the convolutional dilation rate during the expansion process and are unrelated to the shipwreck image. In another embodiment, the preset anti-noise fusion model can be an anti-noise fire suppression strategy. By suppressing the regions unrelated to the target features received by the expanded receptive field, the problem of non-target feature regions being erroneously amplified under large dilation rates is solved, and the localization regions generated by different dilation rates are fused into a complete localization map that highlights the target feature region.

[0036] (2) By fusing the target identification and positioning maps in the target identification and positioning map set, a target feature region positioning map is obtained. In one embodiment, the target identification and positioning maps in the target identification and positioning map set are sorted according to the order of convolution dilation rate from small to large. The target identification and positioning map of the previous level is fused with the target identification and positioning map of the next level step by step to obtain the target feature region positioning map. In one embodiment, the target identification and positioning map with the smallest dilation rate is the area with the most recognizable characteristics of the center of the sunken ship target, and the target identification and positioning map with the larger dilation rate is the low response area such as the surrounding ship hull and outline of the sunken ship target.

[0037] This invention also provides an implementation method for obtaining a multi-scale feature attention-weighted response map, as detailed in (a) to (c) below:

[0038] (a) Obtaining a spatial multi-scale feature map of a shipwreck image. In one embodiment, the shipwreck image is segmented into channels using a segmentation connection unit to obtain a spatial multi-scale feature map.

[0039] (b) Using the weight analysis unit, the attention vectors of the channels at each scale in the spatial multi-scale feature map are extracted to obtain the attention weight information of the channels at each scale. In one implementation, the attention vectors of the channels at different scales are recalibrated and reassigned using the Softmax function to obtain the multi-scale channel attention weights of the new feature map.

[0040] (c) Using attention weight information, the spatial multi-scale feature map is subjected to attention weighting processing to obtain a multi-scale feature attention weighted response map. In one embodiment, the attention weight information of each scale channel is corrected using a preset calibration calculation model to obtain target attention weight information. The target attention weight information is multiplied with the spatial multi-scale feature map of the corresponding scale to obtain an updated target feature map. The target feature maps of each scale are stitched together to obtain a multi-scale feature attention weighted response map. In one embodiment, the weight analysis unit can obtain attention weights from input feature maps of different scales, so that the module can fuse context information of different scales and generate a better pixel-level attention feature map. It can also realize the interaction of attention information and fuse cross-dimensional vectors without destroying the original channel attention vectors, thereby obtaining the entire multi-scale channel attention vector by stitching.

[0041] To facilitate understanding of the target segmentation method for side-scan sonar shipwreck images provided in the above embodiments, this invention provides an application example of the target segmentation method for side-scan sonar shipwreck images. The optimized training module (BHP-Unet) adopts an Encoder-Decoder architecture. The left-hand Encoder performs multi-scale feature extraction through convolution and pooling operations. Specifically, it employs four BHD Module operations to expand the receptive field while maintaining image resolution and spatial hierarchy information. After the HDC operation, it performs three 2*2 Average Pooling operations to reduce computation and memory usage while obtaining multi-scale features. The right-hand Decoder restores resolution through three 2*2 upsampling operations (Upconv) and generates the final feature map through feature concatenation and convolution operations. Before feature concatenation, a PSA Module is introduced to integrate multi-scale spatial information and cross-channel attention into each segmented feature, enabling better information interaction between local and global channel attention. Feature concatenation achieves the fusion of shallow and deep semantic features. The optimization process of the optimized training module is described in [link to optimization process description]. Figure 3 The diagram shows a flowchart of an optimization method for an optimized training module. The method mainly includes the following steps S302 to S304:

[0042] Step S302: The recognition capability of the image segmentation model is improved by using a multi-scale hybrid atrioventricular convolutional (BHD) module. In one implementation, the receptive field is expanded by changing the dilation rate of the convolution kernel at multiple scales, so that low-response target regions can obtain better recognition capabilities by perceiving the surrounding high-response semantics. The knowledge of sparse regions containing recognition features is transferred to adjacent target regions, so that the feature information of the high-response part of the target object can be propagated to adjacent target regions at multiple scales. Finally, the target recognition and localization maps generated under different dilation rates are fused to produce dense and accurate target recognition and localization, thereby fundamentally improving the recognition capability of the segmentation model.

[0043] In one implementation, a small region with the most recognizable features at the center of the shipwreck target can be located using a 3*3 convolutional kernel with an expansion rate of 1. By increasing the expansion rate from 1 to 2 and fusing knowledge from the previous layer, the region near the center of the shipwreck target can be perceived. By further increasing the expansion rate (r=5, 9) and further fusing knowledge from the previous layer, low-response regions such as the surrounding hull and outline of the shipwreck target can be further perceived. This allows for the simultaneous learning of high-response feature regions and the location of more complementary feature regions, thereby comprehensively improving the model's segmentation ability. In another implementation, a noise mitigation strategy can be used to suppress regions unrelated to the target features received by the expanded receptive field. This addresses the problem that some FP (False Positive) non-target feature regions may be incorrectly amplified under large expansion rates, while fusing the localization regions generated by different expansion rates into a complete localization map that highlights the target feature region.

[0044] Step S304: Improve the overall segmentation performance of the image segmentation model by utilizing the attention segmentation model. The attention segmentation model (PSA Module) includes three parts: SPC Module (Split and Concat Module), SEWeight Module, and Fscale. In one implementation, after channel segmentation is performed by the Split and Concat (SPC) module to obtain spatial multi-scale feature maps, the SEWeight module is used to obtain channel attention vectors at different scales. The Softmax function is then used to recalibrate and allocate the channel attention vectors at different scales to obtain new feature map multi-scale channel attention weights. Finally, the element-wise multiplication operation is performed between the new attention weights and the original feature map to obtain the final multi-scale feature attention weighted response map.

[0045] In one implementation, multi-scale feature extraction is achieved in the PSA module through the SPC module. First, the input feature map X is divided into S parts, namely [X0, X1, ... X2].S-1 For each segmented part, there is There are 1 public channel, and the i-th feature map is X. i ∈R C’×H×W Let i = 0, 1, 2, ..., S-1. After the segmentation operation is completed, spatial information on the feature map of each channel is extracted by processing input tensors of multiple scales in parallel. Furthermore, feature maps of different spatial resolutions and depths are generated by using multi-scale convolutional kernels in a pyramid structure. The purpose of the SPC module segmentation is to enable each segmented part to independently learn multi-scale spatial information and establish cross-channel interactions locally. However, as the size of the convolutional kernel increases, the number of parameters in the entire model will increase significantly. Therefore, in order to process input tensors of different scales without increasing computational cost, a grouped convolution method is introduced and applied in parallel to the convolutional kernels. The relationship between the size of the multi-scale convolutional kernel and the size of each group is shown below:

[0046]

[0047] Where K is the kernel size and G is the group size.

[0048] Therefore, the multi-scale feature map function can be obtained as follows:

[0049] F i =Conv(K i ×K i G i (X) i i = 0, 1, 2, ..., S-1

[0050] Where K i =2×(i+1)+1,

[0051] In practical applications, i = 0, 1, 2, 3, F i ∈R C′×H×W The feature maps representing different scales are concatenated to obtain a multi-scale fused feature map, as shown below:

[0052] F = Concat([F0,F1,…,F…) S-1 ])

[0053] Among them, F i ∈R C′×H×W It is the obtained multi-scale feature map, and Concat means concatenating features in the channel dimension.

[0054] After multi-scale feature extraction is completed through the SPC module, the SEWeight module is used to extract channel attention weight information from the multi-scale feature maps. The SEWeight module is divided into two parts: Squeeze and Excitation. The Squeeze part performs one-dimensional compression on the corresponding feature map through global average pooling, that is, compressing the W×H×C' feature map into 1×1×C' as shown below:

[0055]

[0056] Where W×H represents the width and height of the feature map, C is the number of channels, and u i (j,k) represents the element at position (j,k) in the i-th channel, where i∈C';

[0057] After obtaining global features through the Squeeze operation, the relationships between channels are extracted using the Excitation operation, as shown below:

[0058] Ex=σ(g(z,W))=σ(W2δ(W1,z))

[0059] The excitation operation uses the gating mechanism in Sigmoid. By introducing a fully connected layer FC1, the number of channels is reduced to 1 / r of the original number with parameter W1. After activation by the ReLU function δ, the number of channels is restored to the original number by a fully connected layer FC2 with parameter W2. Finally, the weights of each channel are generated by the Sigmoid function σ, where the dimensionality reduction ratio is r = 16.

[0060] The generated weight values ​​are then scaled and applied to the corresponding feature channels F. i In the process, the final output Z is obtained. i As shown below:

[0061] Z i =F scale (u i )=u i ×Sq i

[0062] The SEWeight module is used to extract attention weights from input feature maps at different scales, enabling the module to fuse contextual information at different scales and generate better pixel-level attention feature maps. Furthermore, it allows for the interaction of attention information, fusing cross-dimensional vectors without destroying the original channel attention vectors. Therefore, the entire multi-scale channel attention vector obtained by concatenation is shown below:

[0063]

[0064] in, It's a concat operation, Z i Through F i The obtained attention weight values, Z is a multi-scale attention weight vector.

[0065] To enable each channel to adaptively select different spatial scales, according to Z i Recalibrate the attention vectors for each channel:

[0066]

[0067] The Softmax function is used to obtain the recalibrated multi-scale channel attention weights A. i It contains all spatial location information and attention weights in the channels, thus realizing the interaction between local and global channel attention. The channel attentions of the recalibrated feature maps are then merged again by concatenation, resulting in the attention weights of the entire channel as shown below:

[0068]

[0069] Where A represents the multi-scale channel weights after attention interaction.

[0070] Multiscale channel attention A i Recalibration weights and corresponding scale F i Feature map multiplication:

[0071] Y i =F i ·A i i = 0, 1, 2, ..., S-1

[0072] Where · represents channel multiplication, Y i This represents the feature map after multiplying the multi-scale attention weights through channels. Channel multiplication preserves the representation of the overall features without destroying the information in the original feature map. Finally, the resulting new weighted feature map Y is... i By concatenating the dimensions, we obtain the final output:

[0073] Out = Concat([Y0,Y1,…Y…) S-1 ])

[0074] In summary, the PSA module integrates multi-scale spatial information and cross-channel attention into each segmentation feature group while using a lightweight structure, enabling better information interaction between local and global channel attention and outputting feature maps with multi-scale, global, and long-term information. This achieves the effect of improving the overall segmentation performance of the model while increasing the computational cost slightly.

[0075] In summary, this invention utilizes the analysis results of side-scan sonar shipwreck images using a hybrid dilated convolution model and an attention segmentation model to train the original image segmentation model. This enhances the receptive field while integrating the learning capabilities of deep semantics and shallow features. Furthermore, it establishes long-term dependencies between global and local information while processing multi-scale spatial features, which can significantly reduce image segmentation errors and improve the target segmentation effect of side-scan sonar shipwreck images.

[0076] Regarding the target segmentation method for side-scan sonar shipwreck images provided in the foregoing embodiments, this invention provides a target segmentation device for side-scan sonar shipwreck images. This device is applied to an image segmentation processing system, which is configured with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolutional model and an attention segmentation model. (See [link to relevant documentation]). Figure 4 The diagram shows a target segmentation device for side-scan sonar images of a shipwreck. The device includes the following parts:

[0077] Image acquisition module 402 acquires side-scan sonar images of shipwrecks, wherein the shipwreck images include multiple overlapping shipwrecks.

[0078] The feature region localization module 404 uses a hybrid dilated convolution model to extract multi-scale features from the shipwreck image to obtain a target feature region localization map.

[0079] The attention focus module 406 uses an attention segmentation model to perform attention weight analysis on the shipwreck image and obtain a multi-scale feature attention-weighted response map.

[0080] The segmentation model training module 408 trains the image segmentation model based on the target feature region localization map and the multi-scale feature attention-weighted response map to obtain the target image segmentation model.

[0081] The image segmentation module 410 uses a target image segmentation model to segment the shipwreck image to obtain the target segmented image.

[0082] The data processing apparatus provided in this application embodiment can significantly reduce image segmentation errors and improve the target segmentation effect of side-scan sonar shipwreck images.

[0083] In one embodiment, when performing the step of extracting multi-scale features from the shipwreck image using a hybrid dilated convolution model to obtain a target feature region localization map, the feature region localization module 404 is further configured to: acquire target recognition localization maps corresponding to various convolution dilation rates within a preset dilation rate threshold of the shipwreck image, and establish a target recognition localization map set; and obtain a target feature region localization map by fusing the various target recognition localization maps in the target recognition localization map set.

[0084] In one embodiment, when performing the step of fusing the target identification and positioning maps in the target identification and positioning map set to obtain a target feature region positioning map, the feature region positioning module 404 is further configured to: sort the target identification and positioning maps in the target identification and positioning map set according to the order of convolution dilation rate from smallest to largest; and fuse the previous level target identification and positioning map with the next level target identification and positioning map step by step to obtain a target feature region positioning map.

[0085] In one embodiment, after the step of establishing a target recognition and positioning map set, the feature region positioning module 404 is further used to: use a preset anti-noise fusion model to correct the target feature region positioning map and remove non-key regions in the target recognition and positioning map, wherein the non-key regions are regions that are generated in the target recognition and positioning map corresponding to the convolution dilation rate during the expansion process and are unrelated to the shipwreck image.

[0086] In one embodiment, the attention segmentation model includes a weight analysis unit. When performing the step of using the attention segmentation model to perform attention weight analysis on the shipwreck image to obtain a multi-scale feature attention-weighted response map, the attention concentration module 406 is further configured to: acquire a spatial-level multi-scale feature map of the shipwreck image; use the weight analysis unit to extract the attention vectors of each scale channel in the spatial-level multi-scale feature map to obtain attention weight information for each scale channel; and use the attention weight information to perform attention weighting processing on the spatial-level multi-scale feature map to obtain a multi-scale feature attention-weighted response map.

[0087] In one embodiment, the attention segmentation model includes a segmentation connection unit. When performing the step of acquiring the spatial multi-scale feature map of the shipwreck image, the attention concentration module 406 is further used to: use the segmentation connection unit to perform channel segmentation on the shipwreck image to obtain the spatial multi-scale feature map.

[0088] In one embodiment, when performing the step of using attention weight information to perform attention weighting processing on the spatial multi-scale feature map to obtain a multi-scale feature attention weighted response map, the attention concentration module 406 is further configured to: use a preset calibration calculation model to correct the attention weight information of each scale channel to obtain target attention weight information; multiply the target attention weight information with the spatial multi-scale feature map of the corresponding scale to obtain an updated target feature map; and stitch together the target feature maps of each scale to obtain a multi-scale feature attention weighted response map.

[0089] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0090] This invention provides an electronic device, specifically, the electronic device includes a processor and a storage device; the storage device stores a computer program, and the computer program, when run by the processor, executes the method described in any of the above embodiments.

[0091] Figure 5 The present invention provides a schematic diagram of the structure of an electronic device 100, which includes a processor 50, a memory 51, a bus 52 and a communication interface 53. The processor 50, the communication interface 53 and the memory 51 are connected through the bus 52. The processor 50 is used to execute executable modules, such as computer programs, stored in the memory 51.

[0092] The memory 51 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 53 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc.

[0093] Bus 52 can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0094] The memory 51 is used to store programs. After receiving an execution instruction, the processor 50 executes the programs. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor 50 or implemented by the processor 50.

[0095] Processor 50 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 50 or by instructions in software form. Processor 50 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 51. The processor 50 reads the information in memory 51 and, in conjunction with its hardware, completes the steps of the above method.

[0096] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0097] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0098] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A target segmentation method for side-scan sonar images of a shipwreck, characterized in that, The method is applied to an image segmentation processing system, which is configured with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolutional model and an attention segmentation model. The method includes: Acquire side-scan sonar images of shipwrecks, wherein the shipwreck images include multiple overlapping shipwrecks; Using the hybrid dilated convolution model, multi-scale feature extraction is performed on the shipwreck image to obtain a target feature region localization map; Using the attention segmentation model, attention weight analysis is performed on the shipwreck image to obtain a multi-scale feature attention-weighted response map; The image segmentation model is trained based on the target feature region localization map and the multi-scale feature attention-weighted response map to obtain the target image segmentation model. The target image segmentation model is used to segment the shipwreck image to obtain the target segmented image.

2. The method according to claim 1, characterized in that, The step of using the hybrid dilated convolution model to extract multi-scale features from the shipwreck image to obtain a target feature region localization map includes: Obtain the target recognition and localization map corresponding to each convolutional dilation rate within a preset dilation rate threshold of the shipwreck image, and establish a target recognition and localization map set; The target feature region location map is obtained by fusing the target identification and location maps in the target identification and location map set.

3. The method according to claim 2, characterized in that, The step of fusing the target identification and localization maps in the target identification and localization map set to obtain the target feature region localization map includes: The target recognition and localization maps in the target recognition and localization map set are sorted according to the order of the convolution dilation rate from smallest to largest. The target identification and positioning map described in the previous level is fused with the target identification and positioning map described in the next level step by step to obtain the target feature region positioning map.

4. The method according to claim 2, characterized in that, Following the step of establishing the target identification and localization map set, the following is included: Using a preset noise-resistant fusion model, the target feature region localization map is corrected to remove non-key regions from the target recognition localization map. The non-key regions are regions that are generated in the target recognition localization map corresponding to the convolution dilation rate during the expansion process and are unrelated to the shipwreck image.

5. The method according to claim 1, characterized in that, The attention segmentation model includes a weight analysis unit. The step of using the attention segmentation model to perform attention weight analysis on the shipwreck image to obtain a multi-scale feature attention-weighted response map includes: Obtain the spatial multi-scale feature map of the shipwreck image; Using the weight analysis unit, the attention vectors of the channels at each scale in the spatial multi-scale feature map are extracted to obtain the attention weight information of the channels at each scale. Using the attention weight information, the spatial multi-scale feature map is subjected to attention weighting processing to obtain a multi-scale feature attention-weighted response map.

6. The method according to claim 5, characterized in that, The attention segmentation model includes segmentation connection units, and the step of obtaining the spatial multi-scale feature map of the shipwreck image includes: The segmentation and connection unit is used to perform channel segmentation on the shipwreck image to obtain a spatial multi-scale feature map.

7. The method according to claim 5, characterized in that, The step of using the attention weight information to perform attention weighting processing on the spatial multi-scale feature map to obtain a multi-scale feature attention-weighted response map includes: The attention weight information of each scale channel is corrected using a preset calibration calculation model to obtain the target attention weight information; The target attention weight information is multiplied with the spatial multi-scale feature map at the corresponding scale to obtain the updated target feature map; The target feature maps at each scale are stitched together to obtain a multi-scale feature attention-weighted response map.

8. A target segmentation device for side-scan sonar images of a shipwreck, characterized in that, The apparatus is applied to an image segmentation processing system, which is configured with an optimization training module and an image segmentation model. The optimization training module includes a hybrid dilated convolutional model and an attention segmentation model. The apparatus includes: The image acquisition module acquires side-scan sonar images of shipwrecks, wherein the shipwreck images include multiple overlapping shipwrecks. The feature region localization module uses the hybrid dilated convolution model to perform multi-scale feature extraction on the shipwreck image to obtain a target feature region localization map. The attention focus module uses the attention segmentation model to perform attention weight analysis on the shipwreck image to obtain a multi-scale feature attention-weighted response map. The segmentation model training module trains the image segmentation model based on the target feature region localization map and the multi-scale feature attention-weighted response map to obtain the target image segmentation model. The image segmentation module uses the target image segmentation model to segment the shipwreck image to obtain the target segmented image.

9. A server, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 7.

Images