An underwater target detection method and device based on frequency domain guided feature enhancement

Abstract

The application discloses an underwater target detection method and device based on frequency domain guided feature enhancement, which comprises the following steps: first, obtaining an underwater image to be detected and preprocessing the underwater image to obtain target data; then, inputting the target data into a trained underwater target detection model to extract underwater target information. The model is composed of a backbone network, a neck fusion network and a detection head. The backbone network extracts features in deep and shallow layers. The shallow layer is protected by an information reservation path to protect the high-frequency details of small targets. The deep layer decouples features and refines edges through a frequency domain feature enhancement module. The neck fusion network cross-scale splices and fuses features in deep and shallow layers. The detection head outputs the category and position coordinate information of the target according to the fused features. The application solves the problems of low detection precision and poor stability of the underwater target detection method in the prior art in a complex water environment.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to an underwater target detection method and apparatus based on frequency domain guided feature enhancement. Background Technology

[0002] Intelligent monitoring of marine ranches is a crucial foundation for the sustainable development and precise management of fishery resources. As its core sensing method, the performance of underwater target detection technology directly determines the operational efficiency of autonomous underwater vehicles (AUVs) and fixed monitoring systems. However, in actual complex underwater visual perception tasks, extremely severe natural environmental challenges exist: on the one hand, due to the selective absorption of light by water and strong scattering by suspended particles, underwater images generally exhibit severe color distortion, extremely low physical contrast, and hazy blurring, resulting in a significant weakening of the edge and texture features of targets; on the other hand, the seabed is complex and variable, with benthic organisms such as sea cucumbers and scallops often possessing camouflage colors similar to rocks and seaweed, accompanied by severe semi-burial and dense aggregation of sediment.

[0003] To address the aforementioned challenges, existing underwater target detection technologies mainly offer two types of solutions, but both have significant technical drawbacks: (1) Cascaded architecture based on "enhancement first, detection later": This type of method usually uses physical imaging models (such as Retinex theory) or data-driven generative adversarial networks (GANs) to preprocess underwater images by dehazing or color correction before inputting them into the detector. Drawbacks: Image enhancement and object detection are treated as two independent optimization tasks. The inconsistency of the objective function can easily lead to over-enhancement or the introduction of artificial feature distribution shifts. This forced change in pixel distribution often destroys the high-frequency physical structure in the original image, thus severely limiting the generalization ability of downstream detection models in complex real-world environments.

[0004] (2) End-to-end deep learning detection methods based on the pure spatial domain: Single-stage detection algorithms, represented by the YOLO series, are widely deployed on underwater mobile platforms. To overcome complex background interference, existing technologies have introduced a large number of spatial domain feature fusion and attention mechanisms, such as squeeze and excitation networks (SE) and convolutional block attention modules (CBAM). The drawbacks are: First, most existing methods heavily rely on spatial domain feature extraction. In underwater environments with low contrast and strong scattering, spatial domain convolution struggles to fully extract weak edge and texture information submerged by low-frequency background noise, resulting in a severe decrease in feature representation capabilities. Second, existing lightweight detection networks typically require multiple downsampling operations during deep feature extraction, which can lead to the irreversible loss of extremely limited fine-grained texture features of small targets, resulting in a very high false alarm rate for small targets. Third, due to the lack of effective discrimination of high-frequency background noise in multi-scale features, direct multi-scale feature stitching can easily lead to the cascading amplification of similar high-frequency noise such as seabed reefs and silt, causing serious false alarms and false detections.

[0005] In summary, existing underwater target detection methods cannot simultaneously address target contour extraction under low contrast, feature preservation of small targets, and suppression of complex background noise, ultimately leading to low detection accuracy and poor stability in complex aquatic environments. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide an underwater target detection method and apparatus based on frequency domain guided feature enhancement, which aims to solve the problems of low detection accuracy and poor stability of existing underwater target detection methods in complex aquatic environments.

[0007] The embodiments of the present invention are implemented as follows: An underwater target detection method based on frequency domain guided feature enhancement, the method comprising: The underwater image data to be detected is acquired and preprocessed to obtain the target underwater image data; The underwater image data of the target is input into the underwater target detection model trained by the target detection basic network to obtain the underwater target information contained in the underwater image data; The target detection network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network extracts multi-layer feature information from the image from shallow to deep. The feature fusion network fuses semantic information at different scales. The detection head outputs the target's category information and location coordinate information. The backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, it obtains shallow output features by protecting the high-frequency details of small targets through information preservation paths. After deep feature extraction, it obtains enhanced deep features by frequency domain decoupling and edge refinement based on the frequency domain feature enhancement module. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

[0008] Furthermore, in the above-mentioned underwater target detection method based on frequency domain guided feature enhancement, the step of obtaining shallow output features by protecting the high-frequency detail features of small targets through information preservation paths after shallow feature extraction includes: After shallow feature extraction, 1×1 pointwise convolution is used to compress and reorganize the shallow feature tensor. The recombined features are input into a 3×3 depthwise separable convolution to extract local high-frequency texture information while maintaining spatial resolution. After batch normalization and nonlinear activation, the information-preserving path output features are obtained. ; Introducing learnable control parameters The information retention path output features Compared with the conventional feature output of the backbone network Adaptive weighted fusion is performed to obtain shallow output features. : .

[0009] Furthermore, in the above-mentioned underwater target detection method based on frequency domain guided feature enhancement, the step of obtaining enhanced deep features by frequency domain decoupling and edge refinement of the deep features based on the frequency domain feature enhancement module after deep feature extraction includes: After deep feature extraction, the features are first processed by a 1×1 convolutional block for channel dimensionality reduction and preprocessing, and then split into spatial domain branches and frequency domain branches. In the spatial domain branch, features are processed by a 3×3 depthwise separable convolution and BN+SiLU activation to preserve the original local spatial semantic information in the deep features. In the frequency domain branch, a two-dimensional real fast Fourier transform is used to map the spatial features to the frequency domain to obtain a complex spectrum, and the complex spectrum is decomposed into an amplitude spectrum branch and a phase spectrum branch. An adaptive amplitude residual enhancement mechanism is constructed to dynamically enhance high-frequency components by utilizing the relative relationship between the current frequency amplitude and the global average amplitude. The enhanced amplitude spectrum is recombined with the original phase spectrum and then restored to the spatial domain by a two-dimensional real inverse fast Fourier transform. The spatial features recovered by the frequency domain branch are concatenated with the output features of the spatial domain branch. Finally, a 1×1 convolutional layer and BN+SiLU activation are used to fuse cross-channel features, and the enhanced deep features are output.

[0010] Furthermore, in the aforementioned underwater target detection method based on frequency domain guided feature enhancement, the enhanced amplitude spectrum... The calculation formula is: ; in, For frequency domain spatial coordinates, This represents the global average amplitude of the current feature channel. For learnable frequency domain gain parameters, To prevent division by zero anomalies by extremely small constants.

[0011] Furthermore, in the above-mentioned underwater target detection method based on frequency domain guided feature enhancement, the step of cross-scale splicing and fusion of shallow and deep features in the neck feature fusion network to obtain fused features includes: Shallow and deep features are input into the neck network and concatenated across scales, then baseline features are generated through convolutional blocks. Global average pooling is used to extract channel descriptors containing the global receptive field. ; descriptor Two feature transformation layers are input, with the SiLU activation function applied after the first layer and the Sigmoid activation function applied after the second layer to generate channel weight vectors mapped to the (0,1) interval. ; Channel weight vector Compared with the benchmark features Channel-wise weighted multiplication is performed, followed by spatial detail refinement using 3×3 convolution, and then residual connections are used to connect to baseline features. Adding them together yields a clean fusion feature that filters out high-frequency false alarm noise. : .

[0012] Furthermore, in the above-mentioned underwater target detection method based on frequency domain guided feature enhancement, the formula for calculating the descriptor is: ; Channel weight vector The expression is: ; in, As the baseline feature, H For feature height, W For feature width, c For channel indexing, For spatial location index, This represents the Sigmoid function. Represents the SiLU function. Z This is the channel description vector.

[0013] Furthermore, in the above-mentioned underwater target detection method based on frequency domain guided feature enhancement, the step of the detection head performing target detection based on fused features includes: The fused features are fed into the detection head, and after passing through the classification branch and the bounding box regression branch, the target's category confidence and bounding box coordinate information are calculated.

[0014] Another object of the present invention is to provide an underwater target detection device based on frequency domain guided feature enhancement, the device comprising: The acquisition module is used to acquire underwater image data to be detected and to preprocess the underwater image data to be detected to obtain target underwater image data; The detection module is used to input the underwater image data of the target into the underwater target detection model trained by the target detection basic network, so as to obtain the underwater target information contained in the underwater image data. The target detection network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network extracts multi-layer feature information from the image from shallow to deep. The feature fusion network fuses semantic information at different scales. The detection head outputs the target's category information and location coordinate information. The backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, it obtains shallow output features by protecting the high-frequency details of small targets through information preservation paths. After deep feature extraction, it obtains enhanced deep features by frequency domain decoupling and edge refinement based on the frequency domain feature enhancement module. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

[0015] Another object of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0016] Another object of the present invention is to provide an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method described above.

[0017] This invention performs shallow and deep feature extraction separately through a backbone network. After shallow feature extraction, high-frequency detail features of small targets are preserved through an information retention path to obtain shallow output features, successfully overcoming the downsampling information loss region of deep networks. This effectively preserves and transmits the fine-grained texture features of small targets to subsequent networks, resulting in a substantial improvement in the recall rate and localization accuracy of extremely small targets. After deep feature extraction, a frequency domain feature enhancement module performs frequency domain decoupling and edge refinement on the deep features to obtain enhanced deep features. This fundamentally strengthens the physical contours of targets submerged in low-frequency scattered light, significantly improving the detection accuracy of the model under stringent localization standards and effectively overcoming the problem of missed detections in blurred scenes. Shallow and deep features are fused across scales in a neck feature fusion network to obtain fused features, and the detection head performs target detection based on these fused features. This solves the problems of low detection accuracy and poor stability in complex aquatic environments in existing technologies.

[0018] In addition, the present invention has at least the following beneficial effects: 1. Utilize the relative relationship between the current frequency amplitude and the global average amplitude to dynamically amplify high-frequency structural signals through an adaptive amplitude residual mechanism.

[0019] 2. Addressing the issue that benthic organisms such as sea cucumbers and scallops are extremely small in size and have very few pixels, making them prone to feature loss during downsampling in deep networks, this invention innovatively constructs a small target feature preservation mechanism in the shallow layers of the backbone network. By establishing parallel high-resolution feature extraction and transfer paths and introducing learnable control parameters during fusion, the mechanism successfully overcomes the downsampling information loss region of deep networks, effectively preserving and transferring the fine-grained texture features of tiny targets to subsequent networks. This results in a substantial improvement in the recall rate and localization accuracy of extremely small targets.

[0020] 3. To address the shortcomings of existing networks in easily amplifying high-frequency noise such as seabed reefs, silt, and seaweed during the multi-scale feature fusion stage, this invention designs an Adaptive Feature Fusion (ASRF) module after the splicing nodes. This module utilizes global channel descriptors and attention routing mechanisms to dynamically filter channel-level weights on the fused features, constructing a powerful background noise interception barrier. This not only effectively counteracts the potential background activation side effects of frequency domain enhancement but also enables the model to accurately separate the target from complex background noise, achieving a significant leap in precision while maintaining high recall.

[0021] 4. The proposed space-frequency collaborative detection architecture maintains extremely high lightweight characteristics while performing feature decoupling and multi-scale optimization. Without significantly increasing the number of model parameters and computational complexity (FLOPs), it can meet the real-time inference requirements of underwater mobile devices (e.g., maintaining an inference speed of 250 FPS). Furthermore, based on underlying frequency domain physical feature calibration, this method effectively avoids overfitting to data from a single water area, maintaining robust sensing performance and generalization robustness even in cross-domain unknown marine environments facing sudden changes in water quality, severe color shifts, and extreme optical distortions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the target detection basic network architecture in an underwater target detection method based on frequency domain guided feature enhancement provided in an embodiment of the present invention; Figure 2 This is a flowchart of the underwater target detection method based on frequency domain guided feature enhancement in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the information retention path processing in the underwater target detection method based on frequency domain guided feature enhancement in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the frequency domain feature enhancement module in the underwater target detection method based on frequency domain guided feature enhancement in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the neck feature fusion network in the underwater target detection method based on frequency domain guided feature enhancement in the first embodiment of the present invention; Figure 6 This is a structural block diagram of the underwater target detection device based on frequency domain guidance feature enhancement in the third embodiment of the present invention.

[0023] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] The following embodiments can all be applied to Figure 1 The target detection network shown includes at least a backbone network, a neck feature fusion network, and a detection head. Specifically, it protects the high-frequency detail features of small targets through the Information Preservation Path (IRP). In the deep feature extraction stage of the target detection network, a Frequency Domain Feature Enhancement (FGDB) module is introduced to decouple and enhance the target contour and edge features in low-contrast underwater images. Multi-scale feature fusion is performed in the neck network, and a high-frequency background noise is suppressed through an Adaptive Scale Routing Fusion (ASRF) module. This module and FGDB form a closed-loop balance, effectively filtering out abnormally amplified background noise such as seaweed and reefs.

[0027] It should be pointed out that, Figure 1 The structure shown does not constitute a limitation on the object detection base network. In other embodiments, the object detection base network may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0028] The following will describe in detail how to improve the accuracy of underwater target detection, with reference to specific embodiments and accompanying drawings.

[0029] Example 1 Please see Figure 2 The figure shows an underwater target detection method based on frequency domain guided feature enhancement in the first embodiment of the present invention, the method including steps S10 to S11.

[0030] Step S10: Acquire the underwater image data to be detected and preprocess the underwater image data to be detected to obtain the target underwater image data.

[0031] The underwater image to be detected can be acquired through an underwater vision sensor, and the image size can be uniformized and normalized to serve as input data for the target detection network. Specifically, the underwater image can be a single frame image or a series of video frames.

[0032] Step S11: Input the underwater image data of the target into the underwater target detection model trained by the target detection basic network to obtain the underwater target information contained in the underwater image data.

[0033] Among them, the underwater target detection model has mastered the inherent logic of identifying the underwater target information contained in the underwater image based on the underwater image of the target. Therefore, by inputting the underwater image data of the target into the underwater target detection model, the category information and location coordinate information of the target in the underwater image can be obtained.

[0034] For example, when training an underwater target detection model, underwater images of targets can be collected as a dataset. The dataset can be divided into training, validation, and test sets according to a certain ratio. The dataset can be input into the target detection basic network for training to obtain the optimal detection model. The test set can be input into the trained optimal detection model to verify the optimal detection model and obtain the final underwater target detection model.

[0035] Specifically, the target detection basic network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network is used to extract multi-layer feature information from the image from shallow to deep. The feature fusion network is used to fuse semantic information at different scales. The detection head is used to output the target's category information and location coordinate information.

[0036] Specifically, the backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, the high-frequency detail features of small targets are preserved through information preservation path to obtain shallow output features. After deep feature extraction, the deep features are decoupled in the frequency domain and the edges are refined based on the frequency domain feature enhancement module to obtain enhanced deep features. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

[0037] More specifically, such as Figure 3 As shown, during continuous downsampling in the shallow layers of the backbone network, an information preservation path is constructed parallel to the conventional feature extraction path. After shallow feature extraction, the shallow feature tensors are first compressed and reassembled using 1×1 pointwise convolution to reduce computational cost. Subsequently, the reassembled features are input into 3×3 depthwise convolution to extract local high-frequency texture information while maintaining spatial resolution. After batch normalization (BN) and SiLU nonlinear activation function, the preserved shallow high-frequency features are obtained. Finally, a control parameter that can be learned during network backpropagation is introduced. ,Will Features of the output of the backbone network's conventional path Adaptive weighted residual fusion is performed to obtain shallow output features: ; like Figure 4As shown, after deep feature extraction, the deep features are decoupled in the frequency domain and refined at the edge based on the frequency domain feature enhancement module to obtain enhanced deep features. In this process, frequency domain guidance feature refinement is introduced at the output position of the deep features in the backbone network. Specifically, the features are first processed by a 1×1 convolution block for channel dimensionality reduction and preprocessing, and then split into spatial domain branches and frequency domain branches. In the spatial domain branch, features are processed by a 3×3 depthwise separable convolution and BN+SiLU activation to preserve the original local spatial semantic information in the deep features. In the frequency domain branch, a two-dimensional real fast Fourier transform is used to map the spatial features to the frequency domain to obtain a complex spectrum, and the complex spectrum is decomposed into an amplitude spectrum branch and a phase spectrum branch. An adaptive amplitude residual enhancement mechanism is constructed to dynamically enhance high-frequency components by utilizing the relative relationship between the current frequency amplitude and the global average amplitude. The enhanced amplitude spectrum is recombined with the original phase spectrum and then restored to the spatial domain by a two-dimensional real inverse fast Fourier transform. The spatial features recovered by the frequency domain branch are concatenated with the output features of the spatial domain branch. Finally, a 1×1 convolutional layer and BN+SiLU activation are used to fuse cross-channel features, and the enhanced deep features are output.

[0038] For example, the enhanced amplitude spectrum The calculation formula is:

[0039] in, For frequency domain spatial coordinates, This represents the global average amplitude of the current feature channel. For learnable frequency domain gain parameters, To prevent division by zero anomalies by extremely small constants.

[0040] Finally, shallow and deep features are fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features. The purified multi-scale fused features are fed into the detection head and passed through the classification branch and bounding box regression branch to calculate the target's category confidence and bounding box coordinate information. Finally, the underwater image is labeled and high-precision detection results are output.

[0041] In summary, the underwater target detection method based on frequency domain guided feature enhancement in the above embodiments of the present invention performs shallow feature extraction and deep feature extraction through a backbone network. After shallow feature extraction, the high-frequency detail features of small targets are preserved through an information retention path to obtain shallow output features, successfully overcoming the downsampling information loss region of the deep network. This effectively preserves and transmits the fine-grained texture features of small targets to subsequent networks, resulting in a substantial improvement in the recall rate and localization accuracy of extremely small targets. After deep feature extraction, the deep features are frequency domain decoupled and edge refined based on the frequency domain feature enhancement module to obtain enhanced deep features. This fundamentally strengthens the physical contours of targets submerged in low-frequency scattered light, significantly improving the detection accuracy of the model under stringent localization standards and effectively overcoming the problem of missed detection in blurred scenes. Shallow and deep features are fused across scales in the neck feature fusion network to obtain fused features, and the detection head performs target detection based on the fused features. This solves the problems of low detection accuracy and poor stability in complex aquatic environments in existing technologies.

[0042] Example 2 The underwater target detection method based on frequency domain guidance feature enhancement in this embodiment differs from the underwater target detection method based on frequency domain guidance feature enhancement in Embodiment 1 in the following ways: like Figure 5 As shown, the step of cross-scale splicing and fusion of shallow and deep features in the neck feature fusion network to obtain fused features includes: Shallow and deep features are input into the neck network and concatenated across scales, then baseline features are generated through convolutional blocks. Global average pooling is used to extract channel descriptors containing the global receptive field. ; descriptor Two feature transformation layers are input, with the SiLU activation function applied after the first layer and the Sigmoid activation function applied after the second layer to generate channel weight vectors mapped to the (0,1) interval. ; Channel weight vector Compared with the benchmark features Channel-wise weighted multiplication is performed, followed by spatial detail refinement using 3×3 convolution, and then residual connections are used to connect to baseline features. Adding them together yields a clean fusion feature that filters out high-frequency false alarm noise. : .

[0043] In this process, shallow features and enhanced deep features are input into the neck network for cross-scale concatenation. Since direct concatenation easily amplifies similar high-frequency background noise such as seabed sediment and reefs, an adaptive feature fusion module (ASRF) is connected in series after the concatenation node. After multi-scale feature concatenation, baseline features are first generated through convolutional blocks. ; Global average pooling (GAP) is used to extract channel descriptors containing the global receptive field. ;

[0044] descriptor The input consists of two feature transformation layers (e.g., two fully connected layers or a 1×1 convolutional layer), with the SiLU activation function applied after the first layer and the Sigmoid activation function applied after the second layer, generating channel weight vectors mapped to the (0,1) interval. : ; in, As the baseline feature, H For feature height, W For feature width, c For channel indexing, For spatial location index, This represents the Sigmoid function. Represents the SiLU function. Z The channel description vector; The channel weight vector Compared with the benchmark features Element-wise channel reweighting is then performed; followed by spatial detail refinement using 3×3 convolution, and residual connections are used to connect to baseline features. Adding them together yields a clean fusion feature that filters out high-frequency false alarm noise. : .

[0045] In summary, the underwater target detection method based on frequency domain guided feature enhancement in the above embodiments of the present invention performs shallow feature extraction and deep feature extraction through a backbone network. After shallow feature extraction, the high-frequency detail features of small targets are preserved through an information retention path to obtain shallow output features, successfully overcoming the downsampling information loss region of the deep network. This effectively preserves and transmits the fine-grained texture features of small targets to subsequent networks, resulting in a substantial improvement in the recall rate and localization accuracy of extremely small targets. After deep feature extraction, the deep features are frequency domain decoupled and edge refined based on the frequency domain feature enhancement module to obtain enhanced deep features. This fundamentally strengthens the physical contours of targets submerged in low-frequency scattered light, significantly improving the detection accuracy of the model under stringent localization standards and effectively overcoming the problem of missed detection in blurred scenes. Shallow and deep features are fused across scales in the neck feature fusion network to obtain fused features, and the detection head performs target detection based on the fused features. This solves the problems of low detection accuracy and poor stability in complex aquatic environments in existing technologies.

[0046] Example 3 Please see Figure 6 The image shows an underwater target detection device based on frequency domain guided feature enhancement proposed in the third embodiment of the present invention. The device includes: The acquisition module 100 is used to acquire underwater image data to be detected and preprocess the underwater image data to be detected to obtain target underwater image data; The detection module 200 is used to input the underwater image data of the target into the underwater target detection model trained by the target detection basic network, so as to obtain the underwater target information contained in the underwater image data. The target detection network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network extracts multi-layer feature information from the image from shallow to deep. The feature fusion network fuses semantic information at different scales. The detection head outputs the target's category information and location coordinate information. The backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, it obtains shallow output features by protecting the high-frequency details of small targets through information preservation paths. After deep feature extraction, it obtains enhanced deep features by frequency domain decoupling and edge refinement based on the frequency domain feature enhancement module. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

[0047] The functions or operation steps implemented by the above modules are largely the same as those in the above method embodiments, and will not be repeated here.

[0048] Example 4 In another aspect, the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of the method described in any one of Embodiments 1 to 2 above.

[0049] Example 5 In another aspect, the present invention provides an electronic device, the electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any one of the methods described in Embodiments 1 to 2 above.

[0050] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0051] Those skilled in the art will understand that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0052] More specific examples (a non-exhaustive list) of computer-readable storage media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable storage media can even be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0053] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0054] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0055] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. An underwater target detection method based on frequency domain guided feature enhancement, characterized in that, The method includes: The underwater image data to be detected is acquired and preprocessed to obtain the target underwater image data; The underwater image data of the target is input into the underwater target detection model trained by the target detection basic network to obtain the underwater target information contained in the underwater image data; The target detection network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network extracts multi-layer feature information from the image from shallow to deep. The feature fusion network fuses semantic information at different scales. The detection head outputs the target's category information and location coordinate information. The backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, it obtains shallow output features by protecting the high-frequency details of small targets through information preservation paths. After deep feature extraction, it obtains enhanced deep features by frequency domain decoupling and edge refinement based on the frequency domain feature enhancement module. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

2. The underwater target detection method based on frequency domain guided feature enhancement according to claim 1, characterized in that, The step of obtaining shallow output features by protecting the high-frequency detail features of small targets through information preservation paths after shallow feature extraction includes: After shallow feature extraction, 1×1 pointwise convolution is used to compress and reorganize the shallow feature tensor. The recombined features are input into a 3×3 depthwise separable convolution to extract local high-frequency texture information while maintaining spatial resolution. After batch normalization and nonlinear activation, the information-preserving path output features are obtained. ; Introducing learnable control parameters The information retention path output features Compared with the conventional feature output of the backbone network Adaptive weighted fusion is performed to obtain shallow output features. : 。 3. The underwater target detection method based on frequency domain guided feature enhancement according to claim 1, characterized in that, The step of obtaining enhanced deep features by frequency domain decoupling and edge refinement of the deep features based on the frequency domain feature enhancement module after deep feature extraction includes: After deep feature extraction, the features are first processed by a 1×1 convolutional block for channel dimensionality reduction and preprocessing, and then split into spatial domain branches and frequency domain branches. In the spatial domain branch, features are processed by a 3×3 depthwise separable convolution and BN+SiLU activation to preserve the original local spatial semantic information in the deep features. In the frequency domain branch, a two-dimensional real fast Fourier transform is used to map the spatial features to the frequency domain to obtain a complex spectrum, and the complex spectrum is decomposed into an amplitude spectrum branch and a phase spectrum branch. An adaptive amplitude residual enhancement mechanism is constructed to dynamically enhance high-frequency components by utilizing the relative relationship between the current frequency amplitude and the global average amplitude. The enhanced amplitude spectrum is recombined with the original phase spectrum and then restored to the spatial domain by a two-dimensional real inverse fast Fourier transform. The spatial features recovered by the frequency domain branch are concatenated with the output features of the spatial domain branch. Finally, a 1×1 convolutional layer and BN+SiLU activation are used to fuse cross-channel features, and the enhanced deep features are output.

4. The underwater target detection method based on frequency domain guided feature enhancement according to claim 3, characterized in that, Enhanced amplitude spectrum The calculation formula is: ； in, For frequency domain spatial coordinates, This represents the global average amplitude of the current feature channel. For learnable frequency domain gain parameters, To prevent division by zero anomalies by extremely small constants.

5. The underwater target detection method based on frequency domain guided feature enhancement according to claim 4, characterized in that, The step of cross-scale splicing and fusing shallow and deep features in the neck feature fusion network to obtain fused features includes: Shallow and deep features are input into the neck network and concatenated across scales, then baseline features are generated through convolutional blocks. Global average pooling is used to extract channel descriptors containing the global receptive field. ; descriptor Two feature transformation layers are input, with the SiLU activation function applied after the first layer and the Sigmoid activation function applied after the second layer to generate channel weight vectors mapped to the (0,1) interval. ; Channel weight vector Compared with the benchmark features Channel-wise weighted multiplication is performed, followed by spatial detail refinement using 3×3 convolution, and then residual connections are used to connect to baseline features. Adding them together yields a clean fusion feature that filters out high-frequency false alarm noise. : 。 6. The underwater target detection method based on frequency domain guided feature enhancement according to claim 5, characterized in that, The formula for calculating descriptors is: ； Channel weight vector The expression is: ； in, As the benchmark feature, H For feature height, W For feature width, c For channel indexing, For spatial location index, This represents the Sigmoid function. Represents the SiLU function. Z This is the channel description vector.

7. The underwater target detection method based on frequency domain guided feature enhancement according to claim 1, characterized in that, The steps of target detection based on fused features by the detection head include: The fused features are fed into the detection head, and after passing through the classification branch and the bounding box regression branch, the target's category confidence and bounding box coordinate information are calculated.

8. An underwater target detection device based on frequency domain guidance feature enhancement, characterized in that, The device includes: The acquisition module is used to acquire underwater image data to be detected and to preprocess the underwater image data to be detected to obtain target underwater image data; The detection module is used to input the underwater image data of the target into the underwater target detection model trained by the target detection basic network, so as to obtain the underwater target information contained in the underwater image data. The target detection network includes at least a backbone network, a neck feature fusion network, and a detection head. The backbone network extracts multi-layer feature information from the image from shallow to deep. The feature fusion network fuses semantic information at different scales. The detection head outputs the target's category information and location coordinate information. The backbone network performs shallow feature extraction and deep feature extraction respectively. After shallow feature extraction, it obtains shallow output features by protecting the high-frequency details of small targets through information preservation paths. After deep feature extraction, it obtains enhanced deep features by frequency domain decoupling and edge refinement based on the frequency domain feature enhancement module. The shallow and deep features are then fused across scales in the neck feature fusion network to obtain fused features. The detection head performs target detection based on the fused features.

9. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method as described in any one of claims 1 to 7.

Images