An underwater target detection encoder, detection system and method

Abstract

The application discloses an underwater target detection encoder, a detection system and a method, the encoder comprising an input projection layer, a global semantic modeling unit, a first-stage feature fusion unit and a second-stage feature fusion unit; wherein the input projection layer is used for channel mapping of multi-scale features output by a backbone network, the global semantic modeling unit establishes a dependency relationship between spatial positions in a feature map through a self-attention mechanism; the first-stage feature fusion unit introduces a frequency domain transformation path in a cross-scale feature fusion process, performs frequency screening on features and generates high-frequency perception features; the second-stage feature fusion unit introduces intermediate features in the first stage in a propagation process and performs splicing and reconstruction processing, forming multi-scale feature representation. Based on the encoder, the detection system and the method are constructed, realizing classification and positioning processing of targets in underwater images.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and underwater image processing technology, specifically to an underwater target detection encoder, detection system, and method. Background Technology

[0002] Underwater target detection has significant applications in marine ecological monitoring, underwater robot navigation, and marine resource exploration. However, due to the absorption and scattering of light by water, underwater images often suffer from low contrast, color distortion, and blurred edges, resulting in a significant degradation of the target's texture information and structural features.

[0003] Existing underwater target detection methods mainly employ convolutional neural networks combined with multi-scale feature pyramid structures. For example, feature pyramid networks achieve multi-scale feature fusion through top-down paths and lateral connections. However, these methods typically use element-wise addition or simple concatenation for feature fusion, making it difficult to fully align shallow details with deep semantic information. Furthermore, some methods enhance image edge information through frequency domain filtering, but frequency domain processing is usually treated as a separate preprocessing step, isolated from the deep feature extraction process. High-frequency information is easily weakened by convolution and pooling operations during propagation in deep networks.

[0004] Therefore, the existing technology has the following problems:

[0005] 1. Insufficient ability to extract underwater target edge and texture information;

[0006] 2. The high-frequency enhancement and feature fusion processes are separated, and high-frequency information is easily attenuated during feature propagation;

[0007] 3. Single-stage multi-scale feature fusion struggles to achieve deep alignment between shallow details and deep semantics;

[0008] 4. The local receptive field of convolutional networks limits their ability to model global semantics in complex underwater scenes.

[0009] In view of this, the present invention proposes an underwater target detection encoder, detection system and method. Summary of the Invention

[0010] The purpose of this invention is to provide an underwater target detection encoder, detection system and method. By embedding a high-frequency sensing module in the multi-scale feature fusion process, the high-frequency information and feature fusion process are fused synchronously, thereby improving the accuracy of underwater target detection.

[0011] In a first aspect, the present invention provides an underwater target detection encoder, comprising:

[0012] The input projection layer is used to perform dimensional mapping on the multi-scale features output by the backbone network, so that the number of feature channels at different levels is unified to the preset hidden dimension.

[0013] The global semantic modeling unit, connected to the output of the input projection layer, is used to perform global relation modeling on the lowest spatial resolution feature map in the multi-scale features. It establishes global dependencies between different spatial locations in the feature map through a self-attention mechanism to obtain semantically enhanced features.

[0014] The first-stage feature fusion unit includes:

[0015] The first cross-scale feature fusion module is used to perform spatial resolution alignment on the features of each scale after the input projection layer is processed, and to perform channel stitching and convolution processing on the aligned features to obtain the first cross-scale fused features.

[0016] The high-frequency sensing module, embedded in the feature fusion path of the first cross-scale feature fusion module, performs discrete cosine transform on the input features aligned by spatial resolution to generate a frequency domain coefficient matrix; filters the frequency domain coefficient matrix based on a preset frequency mask to retain high-frequency components; performs inverse transform on the filtered frequency domain coefficients to generate high-frequency sensing features; and performs weighted summation of the high-frequency sensing features and the first cross-scale fusion features to generate the first-stage fusion features.

[0017] The first bidirectional feature propagation path is used to perform bidirectional feature transfer on the multi-scale features after the first stage fusion. In the bottom-up path, the low-level features are downsampled and then concatenated with the high-level features, and feature transformation is performed through a convolutional unit. In the top-down path, the high-level features are upsampled and then concatenated with the low-level features, and feature transformation is performed through a convolutional unit, outputting the first-stage multi-scale features and the first-stage intermediate propagation features.

[0018] The second-stage feature fusion unit includes:

[0019] The second cross-scale feature fusion module is used to perform channel splicing and convolution processing on the first-stage multi-scale features to obtain the second cross-scale fused features.

[0020] The second bidirectional feature propagation path is used to bidirectionally transfer the multi-scale features after the second stage fusion. During the propagation process, the intermediate propagation features of the first stage are spliced, fused and refined with the propagation features of the second stage, so that the features of different scales are gradually aligned in the process of bottom-up semantic enhancement and top-down detail supplementation, thereby outputting a multi-scale feature representation for the object detection task.

[0021] As a preferred embodiment of the first aspect of the present invention, the input projection layer includes:

[0022] The multi-scale feature includes at least a first spatial resolution feature, a second spatial resolution feature, and a third spatial resolution feature, wherein the first spatial resolution is greater than the second spatial resolution, and the second spatial resolution is greater than the third spatial resolution.

[0023] A 1×1 convolutional layer is set up for each level of features to perform channel compression or channel expansion on the input features, so that the features of each level are uniformly mapped to the preset hidden dimension.

[0024] And a normalization layer connected after the 1×1 convolutional layer, used to stabilize feature distribution and enhance the consistency of feature representation.

[0025] As a preferred embodiment of the first aspect of the present invention, the global semantic modeling unit includes:

[0026] Select the high-level feature map with the lowest spatial resolution and the richest semantic information from the backbone network output, and unfold the high-level feature map into a one-dimensional feature sequence according to the spatial dimension.

[0027] The feature sequence is superimposed with positional encoding to represent the spatial positional information of the feature; the positional encoding is generated using a sine-cosine function.

[0028] The feature sequences with added positional encoding are input into the multi-head self-attention layer. Attention weights are calculated by constructing a query matrix, a key matrix, and a value matrix, thereby establishing the dependency relationship between different spatial locations.

[0029] The attention output is input into a feedforward network for feature transformation. The feedforward network consists of two fully connected layers and an activation function.

[0030] Semantic enhancement features are obtained through residual connections and normalization operations, and then restored to a two-dimensional feature map structure to obtain more complete global semantic information.

[0031] As a preferred embodiment of the first aspect of the present invention, the scale feature fusion corresponding to the first cross-scale feature fusion module and the second cross-scale feature fusion module includes:

[0032] The scale alignment operation is used to adjust features with different spatial resolutions to a uniform resolution scale, wherein: the third spatial resolution features are upsampled using interpolation to improve resolution; and the first spatial resolution features are downsampled using a combination of average pooling and convolution.

[0033] Multi-scale perceptual extraction utilizes multiple parallel convolutional branches to extract local structural information from aligned features. Each convolutional branch employs a kernel size of 3×3, 5×5, and 7×7, respectively, and achieves inter-channel information interaction through pointwise convolution.

[0034] As a preferred embodiment of the first aspect of the present invention, the high-frequency sensing module includes:

[0035] The spatial awareness path is used to extract the spatial high-frequency response from the input features. Specifically, for the preset shallow features, the spatial domain features are converted into frequency domain representations through discrete cosine transform, and low-frequency components in the frequency domain are suppressed and high-frequency components are preserved through frequency masking. Then, the spatial high-frequency response is generated through inverse transform. For features at other levels, spatial attention weights are generated through convolution.

[0036] The channel-aware path is used to extract high-frequency responses of input features. Specifically, for preset shallow features, the high-frequency components in the frequency domain are pooled and statistically analyzed, and channel attention weights are generated through a convolutional network. For features at other levels, the input features are directly pooled and statistically analyzed, and channel attention weights are generated through a convolutional network.

[0037] The high-frequency feature refinement path is used to add the output of the spatial perception path and the output of the channel perception path, and then perform convolution and normalization to obtain the high-frequency perception features.

[0038] The high-frequency sensing features are weighted and fused with the cross-scale fusion features output by the first cross-scale feature fusion module through learnable weights.

[0039] As a preferred technical solution of the first aspect of the present invention, when the first bidirectional feature propagation path performs feature transfer, it takes the feature maps of each scale feature branch before splicing and transformation as the intermediate propagation features of the first stage, and constructs an intermediate propagation feature set.

[0040] As a preferred embodiment of the first aspect of the present invention, both the first bidirectional feature propagation path and the second bidirectional feature propagation path include:

[0041] The feature refinement module refines the propagation features using residual blocks containing convolutional layers, normalization layers, and residual connections.

[0042] The bidirectional feature transfer path involves a bottom-up path that uses convolutional downsampling to transfer information from the first spatial resolution feature level to the third spatial resolution feature level, and a top-down path that uses interpolation upsampling to transfer information from the third spatial resolution feature level to the first spatial resolution feature level.

[0043] The first bidirectional feature propagation path is used to output the intermediate propagation features of the first stage, and the second bidirectional feature propagation path is used to concatenate and fuse the propagation features of the second stage with the intermediate propagation features of the first stage to obtain the final multi-scale feature representation.

[0044] As a preferred technical solution of the first aspect of the present invention, the high-frequency sensing module uses learnable weight parameters for balancing when performing weighted summation with the first cross-scale fusion feature; when the second bidirectional feature propagation path introduces the intermediate propagation feature of the first stage, it uses channel dimension splicing to fuse the intermediate feature of the corresponding level of the first stage with the current propagation feature of the second stage, and then performs feature reconstruction through a convolutional layer.

[0045] In a second aspect, the present invention provides an underwater target detection system for loading the underwater target detection encoder described in the first aspect, comprising:

[0046] The image input module is used to acquire raw underwater images and perform pixel normalization and size preprocessing.

[0047] The backbone network module is used for multi-scale feature extraction of the input image;

[0048] The encoder module is used to encode the multi-scale features output by the backbone network, and the encoder module adopts the underwater target detection encoder as described in the first aspect.

[0049] The target detection head module includes a classification branch and a regression branch, which are used to predict the target category and target bounding box location based on the multi-scale features output by the encoder.

[0050] As a preferred embodiment of the second aspect of the present invention, the target detection head module is used to generate target category prediction results and target bounding box prediction results based on the multi-scale features output by the encoder, so as to obtain the final detection result.

[0051] Thirdly, the present invention provides an underwater target detection method, based on the implementation of the second aspect, comprising the following steps:

[0052] Includes the following steps:

[0053] S1: Feature extraction is performed on the input underwater image through a backbone network to obtain multi-scale feature maps with different spatial resolutions;

[0054] S2: Perform channel mapping on the multi-scale features through the input projection layer, so that the features of each layer are uniformly mapped to the preset hidden dimension;

[0055] S3: Input the highest-level features into the global semantic modeling unit, and establish the global dependency relationship between spatial locations in the feature map through positional encoding and multi-head self-attention mechanism to obtain semantically enhanced features;

[0056] S4: Input the multi-scale features into the first-stage feature fusion unit, perform first-stage cross-scale feature fusion through the first cross-scale feature fusion module to obtain cross-scale fusion features, and extract high-frequency components in the frequency domain through the high-frequency sensing module to generate high-frequency sensing features during the aggregation process, and perform weighted fusion of the high-frequency sensing features and cross-scale fusion features to enhance the target edge and texture information.

[0057] S5: Input the first-stage multi-scale features and the first-stage intermediate propagation features into the second-stage feature fusion unit, perform cross-scale feature fusion again through the second cross-scale feature fusion module, and splice and fuse the first-stage intermediate propagation features and the second-stage propagation features during the second bidirectional feature propagation process. Based on the first-stage intermediate propagation features, perform feature reconstruction constraints on the second-stage propagation features to gradually align shallow detail information with deep semantic information.

[0058] S6: Input the multi-scale features output from the second stage into the target detection head to predict the target category and target bounding box position, thereby obtaining the underwater target detection result.

[0059] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0060] This invention introduces a high-frequency perception module into the first-stage cross-scale feature fusion process, enabling the network to extract and enhance high-frequency structural information in underwater images, thereby improving its ability to express target edge and texture features. Simultaneously, by constructing first-stage and second-stage feature fusion units, cross-scale aggregation and progressive refinement of features at different scales are achieved, effectively fusing shallow detail information and deep semantic information in a unified feature space. Furthermore, through a two-stage progressive propagation structure and a bidirectional feature transfer path, multi-scale features are gradually aligned during bottom-up semantic enhancement and top-down detail supplementation, thereby enhancing the integrity and stability of feature representation. Finally, a global semantic modeling unit is used to establish global dependencies between spatial locations in the feature map, improving the model's overall understanding of complex underwater environments. Through the synergistic effect of these structures, this invention effectively improves target detection accuracy even in underwater images with illumination attenuation, edge blurring, and texture degradation, while maintaining controllable model computational complexity and enhancing the stability and reliability of underwater target detection tasks. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0062] Figure 1 This is a structural diagram of the underwater target detection encoder of the present invention;

[0063] Figure 2 This is the global semantic modeling unit in the underwater target detection encoder of the present invention;

[0064] Figure 3 This is a flowchart of the underwater target detection method of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings.

[0066] Throughout the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions. The described embodiments are only a part of the embodiments of this application, not all of them. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0067] Example 1

[0068] like Figure 1-2 As shown, this embodiment provides an underwater target detection encoder, including an input projection layer, a global semantic modeling unit, a first-stage feature fusion unit, and a second-stage feature fusion unit. These modules are connected sequentially according to the feature flow direction. Multi-scale features output from the backbone network are first input to the input projection layer for dimensional mapping, and then global relation modeling is performed by the global semantic modeling unit. Next, in the first-stage feature fusion unit, cross-scale feature fusion is performed through a first cross-scale feature fusion module, and a high-frequency sensing module is embedded during the aggregation process to extract and fuse high-frequency information. Then, the first-stage multi-scale features and the first-stage intermediate propagation features are output through a first bidirectional feature propagation path. Finally, in the second-stage feature fusion unit, the first-stage output features are aggregated again through a second cross-scale feature fusion module, and the first-stage intermediate propagation features are concatenated, fused, and refined with the second-stage propagation features through a second bidirectional feature propagation path, thereby outputting a multi-scale feature representation for target detection tasks. The details are as follows:

[0069] The input projection layer is used to perform a unified dimensionality mapping on the features output from different scales of the backbone network. The multi-scale features output from the backbone network typically include multiple feature maps with different spatial resolutions, such as first-layer features, second-layer features, and third-layer features, with channel numbers C1, C2, and C3, respectively. In the input projection layer, pointwise convolution mapping is performed on each feature layer. First, the input feature map Fi is fed into a 1×1 convolutional layer for channel compression or expansion, unifying the channel number of all feature layers to a preset hidden dimension Ch. Then, a normalization operation is performed on the convolutional output, using batch normalization or layer normalization to stabilize the feature distribution. Finally, a unified dimensionality projection feature is obtained, ensuring that features of different scales have a consistent channel dimension before entering subsequent modules, thus facilitating cross-scale feature fusion.

[0070] The global semantic modeling unit is used to model global dependencies of the highest-level features.

[0071] The highest-level feature is the lowest spatial resolution feature map obtained through multiple downsampling processes among the multi-scale features output by the backbone network; global relation modeling is performed on the lowest spatial resolution feature map among the multi-scale features.

[0072] It should be noted that convolutional neural networks, due to the limitations of their local receptive fields, struggle to fully capture global semantic information in complex underwater environments, thus limiting the model's ability to understand the overall scene. By introducing a self-attention mechanism, the network can establish dependencies between any two spatial locations in the feature map, thereby expanding its receptive range and improving its overall understanding of complex underwater environments. In practical implementation, firstly, the high-level feature map with the lowest spatial resolution and richest semantic information output from the backbone network is selected. This high-level feature map is then expanded into a one-dimensional feature sequence according to the spatial dimension, that is, the two-dimensional feature map is converted into a feature sequence of length H×W. Positional encoding is superimposed on the feature sequence to represent the spatial location information of the features. The positional encoding is generated using a sine-cosine function. The feature sequence with added positional encoding is input into a multi-head self-attention layer. Attention weights are calculated by constructing a query matrix, a key matrix, and a value matrix, thereby establishing dependencies between different spatial locations. The attention output is input into a feedforward network for feature transformation. The feedforward network consists of two fully connected layers and an activation function. Finally, semantically enhanced features are obtained through residual connections and normalization operations, and the two-dimensional feature map structure is restored. The above processing can expand the feature perception range, enabling the network to obtain more complete global semantic information in complex underwater environments.

[0073] A cross-scale feature fusion module is used to achieve cross-scale aggregation of features across multiple scales. Traditional Feature Pyramid Networks (FPNs) fuse features through top-down paths and lateral connections, but typically employ element-wise addition or simple concatenation, making it difficult to fully align shallow detail information with deep semantic information. In this invention, the first cross-scale feature fusion module in the first-stage feature fusion unit and the second cross-scale feature fusion module in the second-stage feature fusion unit have the same basic structure, both achieving cross-scale feature fusion through the synergistic effect of a scale alignment module and a multi-scale perception module. The difference lies in that the first cross-scale feature fusion module embeds a high-frequency perception module during feature fusion to enhance high-frequency structural information, while the second cross-scale feature fusion module does not embed a high-frequency perception module but instead further aggregates and refines based on the first-stage multi-scale features and their intermediate propagation features. Its structure includes a scale alignment module and a multi-scale perception module, wherein:

[0074] The scale alignment module is used to adjust features at different scales to a uniform resolution. For high-level features with low spatial resolution, interpolation upsampling is used to improve their resolution; for shallow features with high spatial resolution, convolutional downsampling is used to reduce their resolution. The downsampling operation includes a combination of average pooling and convolution. First, average pooling is performed on the input features, and then convolution is used for feature extraction. After the above processing, the features of each layer are adjusted to the same spatial resolution, and the features of each layer are concatenated along the channel dimension to form a fused input feature.

[0075] The multi-scale perception module is used to extract local structural information from different receptive fields. This module includes multiple parallel convolutional branches, each using convolutional kernels of different sizes, such as 3×3, 5×5, and 7×7 kernels. Each convolutional branch extracts local structural information at different scales. The outputs of the convolutional branches are fused with the original input features through residual connections, and inter-channel information exchange is achieved through pointwise convolution, thereby obtaining multi-scale fused features.

[0076] A high-frequency sensing module, embedded within the first cross-scale feature fusion module of the first-stage feature fusion unit, is used to extract high-frequency structural information from underwater images. Due to issues such as low contrast, blurred edges, and texture degradation in underwater images, traditional convolution further weakens high-frequency information during feature propagation. Existing methods typically treat frequency domain processing as a separate preprocessing step, separating it from the deep feature extraction process. This leads to high-frequency information being gradually weakened by convolution and pooling operations during propagation in deep networks. This invention directly extracts high-frequency components in the frequency domain during the feature fusion stage, enabling high-frequency features to be fully utilized in subsequent network layers. Compared to independent preprocessing methods, this more effectively preserves and enhances the edge and texture information of underwater targets. This module includes a spatial sensing path and a channel sensing path, wherein:

[0077] Spatial Awareness Path: For preset shallow features, a discrete cosine transform is first performed on the feature map to convert the spatial domain features into a frequency domain representation. High-frequency region coefficients are preserved through a frequency mask, while low-frequency components are filtered out. An inverse transform is then performed on the frequency domain features to obtain the spatial high-frequency response, which is then element-wise interacted with the input features to enhance the spatial high-frequency information. For features at other levels, since they already contain relatively rich semantic information, spatial attention weights are directly generated through convolution and interacted element-wise with the input features, avoiding excessive frequency domain processing.

[0078] Channel-aware path: For preset shallow features, pooling statistics are performed on the high-frequency response in the frequency domain to obtain channel statistical features, and channel attention weights are generated through a convolutional network; for other layer features, pooling statistics are performed directly on the input features, and channel attention weights are generated through a convolutional network.

[0079] High-frequency feature fusion: The outputs of the spatial perception path and the channel perception path are added together, and then processed by convolution and normalization to obtain high-frequency perception features. Subsequently, the high-frequency perception features are weighted and fused with the fused features output by the first cross-scale feature fusion module using learnable weight parameters, thereby enhancing high-frequency information while avoiding excessive noise amplification. Compared to directly using high-frequency features, the weighted fusion method using learnable weights allows the network to adaptively adjust the contribution of high-frequency information during training, enhancing edge and texture features while preventing the negative impact of high-frequency noise on detection performance.

[0080] A two-stage progressive propagation structure is used to progressively propagate, merge, and enhance the fused features from the first and second stages. Traditional feature pyramids employ single-stage fusion, fusing shallow details and deep semantics through simple element-by-element addition or splicing, which struggles to adequately align semantic differences between features at different scales. This invention uses a two-stage progressive propagation structure, achieving progressive alignment of shallow detail information with deep semantic information through two consecutive feature propagation stages and a bidirectional feature transfer path. Compared to single-stage fusion, the two-stage structure enables features to gradually align during bottom-up semantic enhancement and top-down detail supplementation, thereby enhancing the integrity and stability of feature representation and improving cross-scale detection consistency. This structure includes two consecutive feature propagation stages.

[0081] The first stage of feature propagation involves the extraction of fused features through a residual refinement module. This module consists of multiple residual blocks, each containing a convolutional layer, a normalization layer, and residual connections. The processed features are then propagated across layers via downsampling and upsampling paths. This bidirectional feature propagation includes a bottom-up path and a top-down path. In the bottom-up path, intermediate layer features are downsampled via convolution and concatenated with higher-level features, then refined using the residual module. In the top-down path, intermediate layer features are upsampled via interpolation and concatenated with shallow layer features, then refined using the residual module. This bidirectional propagation in the first stage enables preliminary information exchange between features of different scales, outputting intermediate propagated features for further fusion in the second stage.

[0082] The second-stage feature propagation involves using the multi-scale features from the first stage as input and incorporating intermediate propagation features from the first stage into the second-stage propagation process. Feature concatenation and refinement are then performed again to further integrate semantic and detailed information. Compared to the initial fusion in the first stage, the second stage no longer embeds a high-frequency perception module. Instead, it uses intermediate propagation features from the first stage for deeper cross-scale fusion, allowing for a more complete alignment between shallow details and deep semantics. After the second-stage processing, the final multi-scale feature map is output.

[0083] The final output multi-scale features include shallow detail information, deep semantic information, and frequency-domain enhanced high-frequency structural information. These multi-scale features are input into the target detection head to complete target classification and localization tasks. Through the synergistic effect of the encoders, the network can effectively improve target detection accuracy even in underwater images with illumination attenuation, edge blurring, and texture degradation, while improving the stability and reliability of underwater target detection tasks while keeping the model's computational complexity under control.

[0084] Example 2

[0085] Based on the underwater target detection encoder fused with high-frequency sensing described in Embodiment 1, this embodiment provides a complete underwater target detection system, which performs target recognition and localization on underwater images. The detection system includes an image input module, a backbone network module, an underwater target detection encoder fused with high-frequency sensing, and a target detection head module, wherein the encoder adopts the coding structure described in Embodiment 1.

[0086] Specifically, during operation, raw image data of the underwater environment is first acquired through the image input module. This image data can be acquired by underwater camera equipment, an underwater robot equipped with a camera system, or an underwater monitoring device. The acquired images undergo preprocessing, including image resizing and pixel normalization, to unify the input image to a preset resolution and map pixel values ​​to a preset range, thereby improving the stability of subsequent feature extraction processes.

[0087] The preprocessed image is input to a backbone network module for basic feature extraction. This backbone network module extracts multi-scale visual features from the input image and includes a feature enhancement layer and a multi-path downsampling structure. The feature enhancement layer performs initial feature extraction on the input image through convolution operations to obtain a basic feature map. Subsequently, the feature map is downsampled using the multi-path downsampling structure, employing multiple parallel branches to reduce information loss. Specifically, the multi-path downsampling structure includes convolutional branches, pooling branches, and slicing branches. The convolutional branches reduce spatial resolution and extract local structural information through stride convolution; the pooling branches extract significant structural features through pooling operations; and the slicing branches segment and recombine the feature map using interval sampling to preserve original information. The outputs of multiple branches are concatenated along the channel dimension and then fused through convolution operations to obtain a multi-scale feature representation.

[0088] The multi-scale features output from the backbone network module are then input into the underwater target detection encoder with high-frequency sensing for encoding. During encoding, the features at different scales are first channel-mapped through the input projection layer, unifying the mapping of features at each layer to a preset hidden dimension. Then, the highest-level features are input into the global semantic modeling unit, where dependencies between spatial locations in the feature map are established through positional encoding, multi-head self-attention mechanisms, and a feedforward network, thereby obtaining enhanced features containing global semantic information. Next, the multi-scale features are input into the first-stage feature fusion unit, where first-stage cross-scale feature fusion is performed through the first cross-scale feature fusion module. A high-frequency sensing module is embedded during the aggregation process, extracting high-frequency components by performing frequency domain transformation on at least one layer of features. These high-frequency components are then fused with the aggregated features using learnable weights to enhance edge information and texture structure in the underwater image. Finally, the first-stage multi-scale features and the first-stage intermediate propagation features are input into the second-stage feature fusion unit, where they are further aggregated and fused through the second cross-scale feature fusion module and the second bidirectional feature propagation path, resulting in multi-scale encoded features for the detection task.

[0089] The multi-scale features output by the encoder are then input into the target detection head module for target recognition and localization. The target detection head module is used to generate target category prediction results and target bounding box prediction results, thereby obtaining the final target detection result.

[0090] The system architecture described above enables automatic detection and localization of target objects in underwater images. Compared to traditional target detection methods, this embodiment introduces a high-frequency sensing mechanism and a two-stage progressive propagation structure into the encoder. This allows the network to effectively enhance high-frequency structural information during feature fusion and achieve gradual alignment of multi-scale semantic and detail information, thus maintaining high detection accuracy even in underwater images with blurred edges and texture degradation. This system can be applied to various scenarios such as underwater biometrics, underwater equipment detection, and underwater robot visual navigation.

[0091] Example 3

[0092] like Figure 3 As shown, based on Example 1, this embodiment provides an underwater target detection method, including the following steps:

[0093] S1: Multi-scale feature extraction is performed on the input underwater image using a backbone network to obtain multiple feature levels with different spatial resolutions. The input underwater image is then fed into the backbone network for feature extraction. The backbone network can employ a convolutional neural network structure, progressively extracting visual features from the image through multiple convolutional and downsampling operations to form multiple feature maps with different spatial resolutions. For example, the backbone network can output three or four layers of feature maps at different scales, where shallow feature maps have higher spatial resolution and contain richer detail information, while deep feature maps have lower spatial resolution but contain richer semantic information. The multi-scale features obtained through this step provide the foundation for subsequent feature fusion.

[0094] S2: Channel mapping is performed on the multiple feature levels through the input projection layer, uniformly mapping the features of each level to a preset hidden dimension. After obtaining the multi-scale features output by the backbone network, the feature map of each layer is input to the input projection layer for channel mapping processing. The input projection layer adjusts the number of channels for features of different scales through convolution operations, so that the features of each layer are uniformly mapped to the same hidden dimension. Through this processing, the channel dimension differences between features of different scales can be eliminated, thereby providing a unified feature representation space for subsequent cross-scale feature fusion.

[0095] S3: The highest-level feature map is input into the global semantic modeling unit. Global dependencies between spatial locations in the feature map are established through positional encoding, multi-head self-attention, and a feedforward network, resulting in semantically enhanced features. The feature map with the lowest spatial resolution and richest semantic information among the multi-scale features is input into the global semantic modeling unit. First, positional encoding is generated for the feature map to represent the spatial location information of the features. Then, the correlation between different spatial locations in the feature map is calculated through a multi-head self-attention mechanism, thereby establishing global dependencies. Afterward, the attention output is transformed through a feedforward network, and combined with residual connections and normalization operations to obtain semantically enhanced features. This step expands the network's receptive range, enabling the model to capture global semantic information in the image.

[0096] S4: The multiple feature levels are input into the first-stage feature fusion unit for first-stage cross-scale feature fusion. During this process, a high-frequency sensing module is embedded. This module extracts high-frequency components by performing frequency domain transformation on at least one feature layer, and then fuses these high-frequency components with the aggregated features using learnable weights to enhance the edge and texture information of underwater targets. The first cross-scale feature fusion module maintains consistent spatial resolution across different scales through scale alignment operations and extracts local structural information from different receptive fields using a multi-scale sensing module. Simultaneously, a high-frequency sensing module is embedded during feature fusion to extract high-frequency information by performing frequency domain transformation on the features, thereby obtaining edge and texture features in the underwater image. Subsequently, the high-frequency components are weighted and fused with the fused features using learnable weight parameters, thereby enhancing feature representation capabilities.

[0097] S5: The first-stage multi-scale features and the first-stage intermediate propagation features are input into the second-stage feature fusion unit. The multi-scale features are then fused in the second stage through the second cross-scale feature fusion module and the second bidirectional feature propagation path. The first-stage intermediate propagation features output from the first stage are concatenated and fused with the second-stage propagation features during the second-stage propagation process. During the first-stage feature propagation process, the fused features achieve preliminary information propagation between features of different levels through the first bidirectional feature propagation path. In the second stage, by introducing the first-stage intermediate propagation features, feature concatenation and refinement are performed again to further integrate shallow detail information and deep semantic information, thereby improving the multi-scale feature fusion effect.

[0098] S6: The multi-scale features aggregated in the second stage are input into the target detection head for target classification and localization, yielding underwater target detection results. After obtaining the encoded multi-scale features, they are input into the target detection head for detection task processing. The target detection head includes a classification branch and a regression branch, where the classification branch is used to predict the target category probability, and the regression branch is used to predict the target bounding box position parameters, ultimately obtaining the target detection results in the underwater image, including the target category, target location, and confidence score; achieving automatic detection and localization of target objects in underwater images, thereby improving the accuracy and stability of the underwater target detection task.

[0099] Example 4

[0100] This embodiment trains and validates the underwater target detection encoder fused with high-frequency sensing described in this invention on the publicly available underwater target detection dataset DUO (Dataset for Underwater Object Detection) to evaluate its performance in underwater target detection tasks. The DUO dataset contains various typical underwater target categories, such as starfish, sea urchins, sea cucumbers, and scallops. The image data comes from real underwater environments and has features such as light attenuation, color distortion, and suspended particle interference, which can effectively simulate actual underwater detection scenarios.

[0101] During the training phase, the DUO dataset is first preprocessed by adjusting the original images to a preset input size and normalizing the image pixels to ensure their numerical distribution is within a stable range. Simultaneously, the dataset is divided into training and validation sets according to a preset ratio to ensure that the training and model evaluation processes are independent of each other.

[0102] The preprocessed image is input into a complete detection model consisting of a backbone network, an underwater target detection encoder fused with high-frequency sensing, and a target detection head for training. The encoder structure adopts the high-frequency sensing fused coding structure described in Example 1 and is embedded in the overall detection network to encode and enhance the multi-scale features output by the backbone network.

[0103] During model training, the training parameters were set as follows: the optimizer used an adaptive learning rate optimization algorithm, with an initial learning rate of 0.001, allowing dynamic adjustment within the range of 0.0001 to 0.01; the total number of training epochs was set to 200, with data input in batches of 8 per epoch; for data augmentation, a hybrid data augmentation strategy was used from epoch 4 to epoch 104, including random flipping, color jittering, and mosaic enhancement, to improve the model's adaptability to different underwater environmental conditions. Data augmentation strategies were not used in epochs 1 to 3 and epochs 105 to 200 to ensure the model could stably learn data distribution characteristics in the early training and convergence phases. Furthermore, the weight parameters in the high-frequency perception module were initialized to 0.2 at the start of training, allowing the network to appropriately introduce high-frequency information in the initial stage while avoiding excessive influence of high-frequency noise on feature representation.

[0104] After training, the model's performance was evaluated using a validation set. During evaluation, the model's performance was measured by calculating the mean accuracy (mAP50) for the object detection task. The mAP50 and mAP50-95 metrics were used to evaluate detection performance under different cross-union (CUI) thresholds, respectively. The data are shown in Tables 1 and 2.

[0105] Table 1: Performance comparison results of the method of the present invention and various typical target detection models on the DUO underwater target detection dataset:

[0106]

[0107] It can be seen from the above table:

[0108] (1) This invention achieves 83.8% accuracy in the mAP50 metric, a 4.2% improvement compared to the second-best performing DEIM-n (79.6%); and 65.1% accuracy in the more stringent mAP50-95 metric, a 6.1% improvement compared to DEIM-n (59.0%). This significant improvement in accuracy stems from the high-frequency sensing module introduced into the encoder in this invention. Due to the characteristics of blurred edges and degraded textures in underwater images, traditional detectors struggle to clearly identify target contours. This invention, by directly extracting high-frequency components in the frequency domain during the feature fusion stage, can compensate for the edge blurring problem in underwater images, enabling the network to more clearly identify target boundaries, thus achieving a greater improvement in the more stringent mAP50-95 metric (6.1% vs 4.2%), demonstrating that the high-frequency sensing module makes a comprehensive contribution to improving detection accuracy.

[0109] Compared to YOLOv11n, which is also a lightweight detector, the present invention improves mAP50 by 6.9% and mAP50-95 by 9.4%. This comparison further verifies the effectiveness of the underwater target detection encoder with high-frequency sensing fusion described in this invention in underwater scenarios. YOLOv11n adopts a traditional feature pyramid structure, performing multi-scale feature fusion through simple element-by-element addition. In contrast, this invention achieves deep alignment between shallow details and deep semantics through a first-stage feature fusion unit, a second-stage feature fusion unit, and a two-stage progressive propagation structure, enabling more thorough fusion of multi-scale features, especially exhibiting stronger robustness when processing multi-scale underwater targets.

[0110] (2) The present invention has 8.8M parameters and a computational cost of 19.8 GFLOPs. Compared with YOLOv3-tiny (12.1M, 18.9 GFLOPs), which has a similar number of parameters, the present invention improves mAP50 by 18.1% and mAP50-95 by 25.6% with fewer parameters, demonstrating the significant advantage of the present invention in model capacity utilization efficiency. This shows that the present invention, through its carefully designed encoder structure, can achieve higher detection accuracy with fewer parameters, making full use of the expressive power of each parameter.

[0111] Although the computational cost of this invention is higher than that of some lightweight models (such as YOLOv5n's 7.1 GFLOPs), its accuracy gain far outweighs the increase in computational cost. Compared to YOLOv5n, this invention improves mAP50 by 24.7% and mAP50-95 by 27.3%, while increasing computational cost by only 12.7 GFLOPs. This demonstrates that the high-frequency sensing mechanism and the two-stage progressive propagation structure can achieve a significant accuracy improvement with relatively low additional computational cost, reflecting the advantage of this invention in balancing accuracy and efficiency.

[0112] Table 2: Ablation experiment results of each key module of the model of the present invention;

[0113]

[0114] To verify the effectiveness of each module, a traditional feature pyramid encoder with a PAN+FPN structure was used as the baseline model. Ablation comparisons were then performed by removing the high-frequency sensing module from the complete model of this invention. All three experiments used the same backbone network and target detection head. As shown in the table above:

[0115] (1) Comparing the "without high-frequency sensing module" (80.4% mAP50) and the "invention" (83.8% mAP50), the introduction of the high-frequency sensing module improved mAP50 by 3.4% and mAP50-95 by 4.2%. This improvement is due to the fact that the high-frequency components in the frequency domain of underwater images contain the edge and texture information of the target, which is crucial for accurately locating the target boundary. The invention extracts high-frequency coefficients through discrete cosine transform and fuses them through learnable weights, enabling the network to directly enhance these key information during the feature fusion stage. Further qualitative analysis shows that the detection accuracy of blurred targets decreases by 5%-15% after removing the high-frequency sensing module, which fully demonstrates that the high-frequency sensing module plays a key role in restoring degraded edge and texture information in underwater images. Although the introduction of the high-frequency sensing module increases the computation by 8.4 GFLOPs, the resulting improvement in accuracy is significant, reflecting the effectiveness and necessity of the high-frequency sensing mechanism.

[0116] The effect of cross-scale feature fusion and two-stage progressive propagation: Compared with the "traditional feature pyramid encoder" (using PAN+FPN structure, 79.6% mAP50) and the "without high-frequency sensing module" (using two-stage cross-scale feature fusion and two-stage progressive propagation, 80.4% mAP50), the first-stage feature fusion unit, the second-stage feature fusion unit, and the two-stage progressive propagation structure proposed in this invention improve mAP50 by 0.8 percentage points and mAP50-95 by 1.9%. Although the computational cost increases from 7.0 GFLOPs to 11.4 GFLOPs, the parallel convolution of the multi-scale sensing module and the bidirectional feature propagation path effectively enhance the cross-scale feature fusion capability.

[0117] The mechanism behind this improvement lies in the fact that traditional FPN uses a single-stage fusion method, fusing features at different scales through simple element-wise addition, which makes it difficult to fully align shallow details with deep semantics. The dual-stage progressive propagation structure of this invention uses two consecutive feature propagation stages to gradually align shallow detail information with deep semantic information. The first stage performs initial fusion through bidirectional feature transfer, and the second stage retains the feature information from the first stage through residual connections and performs deeper refinement, resulting in more complete fusion of multi-scale features. Further comparison between the "traditional feature pyramid encoder" and the "invention" (83.8% mAP50) shows that the complete dual-stage progressive propagation structure improves mAP50 by 4.2% and mAP50-95 by 6.1%. Qualitative analysis shows that removing the dual-stage progressive propagation reduces the detection consistency of multi-scale targets by 3%-10%, indicating that the dual-stage structure effectively achieves deep fusion of shallow details and deep semantics through the gradual alignment of feature distribution and bidirectional propagation paths, significantly improving cross-scale detection capabilities.

[0118] (2) With a parameter count of 8.8M and a computational cost of 19.8 GFLOPs, the complete model of this invention, compared to the "traditional feature pyramid encoder" (parameter count of 5.3M and computational cost of 7.0 GFLOPs), although the parameter count increased by 3.5M and the computational cost increased by 12.8 GFLOPs, achieved a 4.2% improvement in mAP50 and a 6.1% improvement in mAP50-95. This indicates that the synergistic effect of the high-frequency perception module, the first-stage feature fusion unit, the second-stage feature fusion unit, the two-stage progressive propagation structure, and the global semantic modeling unit enables this invention to achieve a significant improvement in accuracy with reasonable computational overhead.

[0119] Specifically, the high-frequency perception module enhances edge and texture information, providing a clearer feature representation for the first-stage cross-scale feature fusion. The first-stage feature fusion unit achieves preliminary fusion and high-frequency enhancement of the original multi-scale features through scale alignment and multi-scale perception. The second-stage feature fusion unit introduces intermediate propagation features from the first stage to achieve deeper cross-scale information alignment and feature refinement. The two-stage progressive propagation structure enables shallow details and deep semantics to fully interact through gradual alignment in two stages. The global semantic modeling unit uses a self-attention mechanism to enable the network to capture global contextual information. These four modules work together to form a complete feature encoding and enhancement system, enabling the invention to achieve a good balance between accuracy and efficiency in underwater target detection tasks.

[0120] In summary, the comparative analysis of Tables 1 and 2 shows that the underwater target detection encoder with high-frequency sensing fusion proposed in this invention can significantly improve the accuracy of underwater target detection. Compared with the feature fusion structure without high-frequency sensing mechanism, the model's detection accuracy for targets with blurred edges and degraded textures is significantly improved after introducing the high-frequency sensing module. Simultaneously, the two-stage progressive propagation structure achieves gradual alignment between multi-scale features, enabling effective fusion of shallow detail information and deep semantic information, thereby further improving the model's detection stability in complex underwater environments. This invention achieves a significant accuracy improvement with relatively low additional computational cost through the high-frequency sensing module and the two-stage progressive propagation structure, demonstrating its advantage in balancing accuracy and efficiency. Experimental verification in this embodiment proves that the underwater target detection encoder with high-frequency sensing fusion proposed in this invention has good detection accuracy and stability in underwater target detection tasks, and can significantly improve model performance while maintaining controllable computational complexity.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An underwater target detection encoder, characterized in that, include: The input projection layer is used to perform dimensional mapping on multi-scale features, with a unified number of channels equal to the preset hidden dimension; among them, high-level features are feature maps with the smallest spatial resolution, and low-level features are feature maps with the largest spatial resolution. The global semantic modeling unit, connected to the output of the input projection layer, is used to process high-level features into feature sequences and superimpose positional codes. The sequence positional correlation is calculated through multi-head self-attention, and after transformation by the feedforward network, it is restored into a two-dimensional semantically enhanced feature map. The first-stage feature fusion unit includes: The first cross-scale feature fusion module is used to perform spatial alignment, channel concatenation and convolution processing on multi-scale features to obtain the first cross-scale fused features. The high-frequency sensing module is located inside the first cross-scale feature fusion module and in the feature processing path after spatial resolution alignment and before channel stitching. It is used to perform discrete cosine transform on the aligned input features to generate a frequency domain coefficient matrix, filter high-frequency components based on a preset frequency mask, generate high-frequency sensing features through inverse transformation, and weighted sum with the cross-scale fusion features to generate the first-stage fusion features. The first bidirectional feature propagation path is used to perform bidirectional propagation on the first-stage fused features: the bottom-up path downsamples the low-level features and then concatenates and transforms them with the high-level features; the top-down path upsamples the high-level features and then concatenates and transforms them with the low-level features, outputting the first-stage multi-scale features and the first-stage intermediate propagation features. The second-stage feature fusion unit includes: The second cross-scale feature fusion module is used to perform concatenation and convolution processing on the output of the first stage to obtain the second cross-scale fused features. The second bidirectional feature propagation path is used to introduce the intermediate propagation features from the first stage into the bidirectional propagation for splicing and convolution processing, and output a multi-scale feature representation for target detection.

2. The underwater target detection encoder according to claim 1, characterized in that, The multi-scale features processed by the input projection layer include at least a first spatial resolution feature, a second spatial resolution feature, and a third spatial resolution feature, wherein the first spatial resolution is greater than the second spatial resolution, and the second spatial resolution is greater than the third spatial resolution. The input projection layer includes: a 1×1 convolutional layer set for each level of features, used to perform channel compression or channel expansion on the input features, so that the features of each level are uniformly mapped to a preset hidden dimension; And a normalization layer connected after the 1×1 convolutional layer, used to stabilize feature distribution and enhance the consistency of feature representation.

3. The underwater target detection encoder according to claim 2, characterized in that, The global semantic modeling unit includes: The high-level feature map with the lowest spatial resolution is expanded into a one-dimensional feature sequence according to the spatial dimension; a positional encoding is superimposed on the feature sequence, wherein the positional encoding is generated based on a sine-cosine function; The feature sequences with added positional encoding are input into the multi-head self-attention layer. Attention weights are calculated by constructing a query matrix, a key matrix, and a value matrix, thereby establishing the dependency relationship between different spatial locations. The attention output is input into a feedforward network for feature transformation. The feedforward network consists of two fully connected layers and an activation function. Semantic enhancement features are obtained through residual connections and normalization operations, and then restored to a two-dimensional feature map structure to obtain a more complete semantic enhancement feature map.

4. The underwater target detection encoder according to claim 3, characterized in that, The scale feature fusion corresponding to the first cross-scale feature fusion module and the second cross-scale feature fusion module includes: The scale alignment operation is used to adjust features with different spatial resolutions to a uniform resolution scale, wherein: the third spatial resolution features are upsampled using interpolation to improve resolution; and the first spatial resolution features are downsampled using a combination of average pooling and convolution. Multi-scale perceptual extraction utilizes multiple parallel convolutional branches to extract local structural information from aligned features. Each convolutional branch has a different kernel size to extract local structural information from different receptive fields, and inter-channel information interaction is achieved through pointwise convolution.

5. An underwater target detection encoder according to claim 4, characterized in that, The high-frequency sensing module includes: The spatial awareness path is used to extract the spatial high-frequency response from the input features. Specifically, for the preset shallow features, the spatial domain features are converted into frequency domain representations through discrete cosine transform, and low-frequency components in the frequency domain are suppressed and high-frequency components are preserved through frequency masking. Then, the spatial high-frequency response is generated through inverse transform. For features at other levels, spatial attention weights are generated through convolution. The channel-aware path is used to extract high-frequency responses of input features. Specifically, for preset shallow features, the high-frequency components in the frequency domain are pooled and statistically analyzed, and channel attention weights are generated through a convolutional network. For features at other levels, the input features are directly pooled and statistically analyzed, and channel attention weights are generated through a convolutional network. The high-frequency feature refinement path is used to add the output of the spatial perception path and the output of the channel perception path, and then perform convolution and normalization to obtain the high-frequency perception features. The high-frequency sensing features are weighted and fused with the cross-scale fusion features output by the first cross-scale feature fusion module through learnable weights.

6. An underwater target detection encoder according to claim 5, characterized in that, When performing feature transfer, the first bidirectional feature propagation path uses the feature maps of each scale feature branch before splicing and transformation as the intermediate propagation features of the first stage, and constructs an intermediate propagation feature set.

7. An underwater target detection encoder according to claim 6, characterized in that, Both the first bidirectional feature propagation path and the second bidirectional feature propagation path include: The feature refinement module refines the propagation features using residual blocks containing convolutional layers, normalization layers, and residual connections. The bidirectional feature transfer path involves a bottom-up path that uses convolutional downsampling to transfer information from the first spatial resolution feature level to the third spatial resolution feature level, and a top-down path that uses interpolation upsampling to transfer information from the third spatial resolution feature level to the first spatial resolution feature level. The first bidirectional feature propagation path is used to output the intermediate propagation features of the first stage, and the second bidirectional feature propagation path is used to concatenate and fuse the propagation features of the second stage with the intermediate propagation features of the first stage to obtain the final multi-scale feature representation.

8. An underwater target detection encoder according to claim 7, characterized in that, When the high-frequency sensing module performs a weighted summation with the first cross-scale fusion feature, it uses learnable weight parameters for balancing. When the second bidirectional feature propagation path introduces the intermediate propagation feature of the first stage, it uses a channel dimension splicing method to fuse the intermediate feature of the corresponding level of the first stage with the current propagation feature of the second stage, and then performs feature reconstruction through a convolutional layer.

9. An underwater target detection system, characterized in that, include: The image input module is used to acquire raw underwater images and perform pixel normalization and size preprocessing. The backbone network module is used for multi-scale feature extraction of the input image; The encoder module is used to encode the multi-scale features output by the backbone network, and the encoder module adopts an underwater target detection encoder as described in any one of claims 1-8. The target detection head module includes a classification branch and a regression branch, which are used to predict the target category and target bounding box location based on the multi-scale features output by the encoder.

10. An underwater target detection method, based on the implementation of the underwater target detection system of claim 9, characterized in that, Includes the following steps: S1: Extract features from the input underwater image using a backbone network to obtain multi-scale feature maps with different spatial resolutions; S2: Perform channel mapping on the multi-scale features through the input projection layer, so that the features of each layer are uniformly mapped to the preset hidden dimension; S3: Input the highest-level features into the global semantic modeling unit, and establish the global dependency relationship between spatial locations in the feature map through positional encoding and multi-head self-attention mechanism to obtain semantically enhanced features; S4: Input the multi-scale features into the first-stage feature fusion unit, perform first-stage cross-scale feature fusion through the first cross-scale feature fusion module to obtain cross-scale fusion features, and extract high-frequency components in the frequency domain through the high-frequency sensing module to generate high-frequency sensing features during the aggregation process, and perform weighted fusion of the high-frequency sensing features and cross-scale fusion features to enhance the target edge and texture information. S5: Input the first-stage multi-scale features and the first-stage intermediate propagation features into the second-stage feature fusion unit, perform cross-scale feature fusion again through the second cross-scale feature fusion module, and splice and fuse the first-stage intermediate propagation features and the second-stage propagation features during the second bidirectional feature propagation process. Based on the first-stage intermediate propagation features, perform feature reconstruction constraints on the second-stage propagation features to gradually align shallow detail information with deep semantic information. S6: Input the multi-scale features output from the second stage into the target detection head to predict the target category and target bounding box position, thereby obtaining the underwater target detection result.

Images