Target detection method combined with underwater image enhancement

Abstract

The invention provides a target detection method combined with underwater image enhancement, and the method comprises the steps: S1, carrying out the enhancement processing of an underwater original image through UnitModule, and generating an enhanced image with a clearer object boundary; s2, respectively extracting features of the underwater original image and the enhanced image by using a backbone network; s3, inputting the extracted underwater original image features and enhanced image features into a dual-channel feature cross enhancement module, and performing feature cross enhancement; s4, fusing the features of the two branches after cross enhancement through a double-branch three-dimensional adaptive feature fusion module, and filtering noise generated during image enhancement; and S5, after passing through the neck network, entering a detection head to output a detection result. According to the invention, through the underwater image enhancement module, the dual-channel feature cross enhancement module and the dual-branch three-dimensional adaptive feature fusion module, the features of the enhanced image and the original image are fused, and the target detection effect under the condition of poor underwater image quality is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of underwater image target detection technology, and more particularly to a target detection method that combines underwater image enhancement. Background Technology

[0002] In underwater environments, images suffer severe degradation due to unique optical properties such as light absorption and scattering. This degradation manifests as color deviation, low contrast, and uneven brightness, with varying degrees of these problems in different underwater regions. Underwater image enhancement methods aim to restore the visual quality of these degraded images, supporting subsequent underwater target detection tasks.

[0003] Currently, most algorithms in underwater target detection tasks employ underwater image enhancement as a preprocessing step to improve detection accuracy. In addition, there are some end-to-end underwater image enhancement and underwater target detection architectures, which typically use only the enhanced image for target detection. Meanwhile, some research attempts to fuse original image features and enhanced image features to further improve detection performance.

[0004] However, existing technologies have some shortcomings. On the one hand, algorithms that use only enhanced images for target detection, while providing clearer object boundaries, lack the natural scene features provided by the original image, which to some extent limits the effectiveness of underwater target detection. On the other hand, current fusion algorithms have deficiencies in the cross-fusion of original and enhanced image features and noise filtering, resulting in unsatisfactory target detection performance even with poor underwater image quality. Therefore, how to effectively fuse original and enhanced image features while filtering out noise has become a key issue in improving the effectiveness of underwater target detection. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a target detection method combining underwater image enhancement. The invention obtains an enhanced image beneficial for underwater target detection through an underwater image enhancement module; and fuses the features of the enhanced image and the original image through a dual-path feature cross-enhancement module and a dual-branch 3D adaptive feature fusion module. The enhanced image helps provide clearer object boundaries, while the original image retains the natural features of the scene, thereby improving the target detection performance even under conditions of poor underwater image quality.

[0006] The technical means employed in this invention are as follows: A target detection method combining underwater image enhancement includes: S1. Enhance the original underwater image using UnitModule to generate an enhanced image with clearer object boundaries; S2. Use a backbone network to extract features from the original underwater image and the enhanced image, respectively; S3. Input the extracted underwater raw image features and enhanced image features into the dual-path feature cross-enhancement module to perform feature cross-enhancement; S4. The features of the two branches after cross-enhancement are fused through the dual-branch 3D adaptive feature fusion module to filter out noise generated during image enhancement; S5. After passing through the neck network, the detection head outputs the detection results.

[0007] Furthermore, the UnitModule enhances the image based on the inverse process of the improved Koschmieder model, specifically as follows: The original underwater image is input into the UnitModule as a degraded image to learn the transmission map and the underwater global background light; Based on the learned transmission map and underwater global background light, the enhanced image is calculated using the following formula:

[0008] in, This indicates an undegraded image, i.e., an enhanced image; This indicates a degraded image, i.e., the original underwater image; This represents the transmission diagram, which physically indicates the proportion of light that reaches the visual sensor after particle attenuation. This represents the underwater global background light, which is the average value of each channel in the input image.

[0009] Furthermore, the unsupervised learning strategy of the UnitModule is as follows: Input a degraded image, denoted as For degraded images Set a hyperparameter The image was calculated to obtain further degraded images. The calculation formula is as follows:

[0010] By minimizing and The difference between them yields the transport graph loss, as follows:

[0011] in, Represents the unsupervised loss of the transport graph; Represents pixels; Indicates based on degraded image The estimated value of UnitModule; Indicates based on degraded image Further degraded images The estimated value of UnitModule; Indicates control The hyperparameters for the degree of degradation range from 0 to 1. The key to achieving transport graph loss is to pay attention to the image... and images Underwater global background light Exactly the same.

[0012] Further, step S2 includes: S21. Extract features from the original underwater image using a backbone network, denoted as... , ,in, Represents the feature layer; The number of channels representing the features of the original underwater image; Indicates the height of features in the original underwater image; Indicates the width of the features in the original underwater image; S22. Extract and enhance the features of the image using a backbone network, denoted as... , ,in, The number of channels representing the enhanced image features; Indicates the height of the enhanced image features; This indicates the width of the enhanced image features.

[0013] Further, step S3 includes: S31. Extracting underwater raw image features from the backbone network and enhance image features The input image serves as the input image for the dual-path feature cross-enhancement module; S32. Reshape the input image using features. The reshaped features are as follows: and The input image is divided into non-overlapping segments. A partial window, in which and All represent the total number of windows. Indicates window size; S33. Introduce a feature cross-enhancement mechanism, the formula of which is as follows:

[0014]

[0015] The query for the original image features is multiplied by the key matrix of the enhanced image features, and then a relative positional bias is added. The attention score is used to dynamically adjust the attention score, thereby reflecting the relative positional relationship between elements in the input sequence; then, softmax is used to normalize the result, generating the attention score, which is then multiplied by the value matrices of the original image features and the enhanced image features, respectively; in the above formula, Represents the query, key, and value mapped by features; Features of the original underwater image Queries, keys, and values; Enhance image features respectively Queries, keys, and values; S34, calculated WCA Features of the original image and features of the enhanced image The output features of the layer are as follows:

[0016] in, and They represent Features of the original image and features of the enhanced image The output features of the layer; Indicates normalization; S35, in WCA Based on this, the residual block output features of the original image features and the MLP module are calculated. The layer output features are as follows:

[0017]

[0018] in, and The residual block output features representing the original image features and the MLP module's features are respectively... Layer output features; S36, In WCA Based on this, the residual block output features of the enhanced image features and the MLP module are calculated. The layer output features are as follows:

[0019]

[0020] in, and These represent the residual block output features for enhancing image features and the MLP module, respectively. Layer output features.

[0021] Further, step S4 includes: S41. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the width dimension, as follows:

[0022] in, This represents the feature layer to be encoded in the dual-branch 3D adaptive feature fusion module. In the In terms of width The value at that location, This represents the feature layer obtained after encoding along the width dimension. exist The value at that location, These represent the batch number, channel number, height, and width of the feature layer, respectively. Represents the weight values ​​during the encoding process; S42. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the height dimension, as follows:

[0023] in, This represents the feature layer obtained after encoding along the height dimension. exist The value at that location,

[0024] S43. The adaptive weights encoded along the width and height dimensions are calculated using the activation function, as shown in the following formula:

[0025] in, Representation of features and Adaptive weights in the height dimension; Representation of features and Adaptive weighting in the width dimension; Representing dimensions and dimensions Exchange operations between them; Representing dimensions and dimensions Exchange operations between them; Representation of features and ; S44. Copy the adaptive weights encoded along the width dimension along the width direction, and copy the adaptive weights encoded along the height dimension along the height direction, so that their shape becomes... ; S45. The fused features are calculated using the weights encoded along the width and height dimensions, respectively. The calculation formula is as follows:

[0026] in, and They represent and The fused features are obtained by fusing along the height and width dimensions; Represents element-wise multiplication. , These are features from the original image and the enhanced image, respectively; S46. Calculate the fusion features using the activation function. and The adaptive weights encoded along the width and height dimensions are formulated as follows:

[0027]

[0028] S47, Based on fusion features and Adaptive weights encoded along the width and height dimensions, adaptive fusion features and The final output features are calculated as follows:

[0029] in, The final output features after adaptive fusion .

[0030] Furthermore, the method of the present invention also includes: The steps for training the model using UnitModule unsupervised loss and detection loss as the total loss function are as follows:

[0031] in, Indicates detection loss, This indicates unsupervised loss. This represents the weighting coefficient.

[0032] Compared with the prior art, the present invention has the following advantages: 1. The present invention provides a target detection method that combines underwater image enhancement. The method enhances the original underwater image using UnitModule to generate an enhanced image with clearer object boundaries, providing a higher quality image foundation for subsequent target detection and helping to improve the accuracy of target detection.

[0033] 2. The present invention provides a target detection method that combines underwater image enhancement. It utilizes a dual-path feature cross-enhancement module to cross-enhance the features of the enhanced image and the original image, making full use of the advantages of both, thereby enhancing the accuracy and reliability of target detection and enabling the model to better identify target objects.

[0034] 3. The target detection method combined with underwater image enhancement provided by this invention adopts a dual-branch three-dimensional adaptive feature fusion module to fuse the cross-enhanced features and filter the noise generated during image enhancement, thereby further improving the quality of features, reducing false detections and false negatives, and enhancing the robustness of target detection.

[0035] 4. The present invention provides a target detection method that combines underwater image enhancement. It uses UnitModule unsupervised loss and detection loss to train the model, so that the enhanced image is conducive to target detection, rather than simply pursuing visual enhancement effect, thereby improving the detection performance of the model in practical applications.

[0036] Based on the above reasons, this invention can be widely applied in fields such as underwater image target detection. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of the target detection algorithm combining underwater image enhancement according to the present invention.

[0039] Figure 2 This is a flowchart of the UnitModule of the present invention.

[0040] Figure 3 This is the unsupervised loss graph for the UnitModule of this invention.

[0041] Figure 4 This is a structural diagram of the dual-path feature cross-enhancement module of the present invention.

[0042] Figure 5 This is a structural diagram of the dual-branch 3D adaptive feature fusion module of the present invention. Figure 6 The image shows a visual comparison of the method of the present invention provided in the embodiments of the present invention. Detailed Implementation

[0043] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0044] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification, claims and accompanying drawings of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such processes, methods, products or devices.

[0045] like Figure 1 As shown, the present invention provides a target detection method combining underwater image enhancement, comprising: S1. Enhance the original underwater image using UnitModule to generate an enhanced image with clear object boundaries; S2. Use a backbone network to extract features from the original underwater image and the enhanced image, respectively; S3. Input the extracted underwater raw image features and enhanced image features into the dual-path feature cross-enhancement module to perform feature cross-enhancement; S4. The features of the two branches after cross-enhancement are fused by the dual-branch 3D adaptive feature fusion module to filter out noise generated during image enhancement; S5. After passing through the neck network, the detection head outputs the detection results.

[0046] In specific implementation, as a preferred embodiment of the present invention, such as Figure 2 As shown, the UnitModule enhances the image based on the inverse process of the improved Koschmieder model, specifically as follows: The original underwater image is input into the UnitModule as a degraded image to learn the transport map and the underwater global background light; Based on the learned transmission map and underwater global background light, the enhanced image is calculated using the following formula:

[0047] in, This indicates an undegraded image, i.e., an enhanced image; This indicates a degraded image, i.e., the original underwater image; This represents the transmission diagram, which physically indicates the proportion of light that reaches the visual sensor after particle attenuation. This represents the underwater global background light, which is the average value of each channel in the input image.

[0048] In specific implementation, as a preferred embodiment of the present invention, such as Figure 3 As shown, the unsupervised learning strategy of the UnitModule is as follows: Input a degraded image, denoted as For degraded images Set a hyperparameter The image was calculated to obtain further degraded images. The calculation formula is as follows:

[0049] By minimizing and The difference between them yields the transport graph loss, as follows:

[0050] in, Represents the unsupervised loss of the transport graph; Represents pixels; Indicates based on degraded image The estimated value of UnitModule; Indicates based on degraded image Further degraded images The estimated value of UnitModule; Indicates control The hyperparameters for the degree of degradation range from 0 to 1. The key to achieving transport graph loss is to pay attention to the image... and images Underwater global background light Exactly the same. In this embodiment, the unsupervised learning strategy of UnitModule only needs to consider how to learn the transport graph. The detector can be trained without any additional dataset.

[0051] In a specific implementation, as a preferred embodiment of the present invention, step S2 includes: S21. Extract features from the original underwater image using a backbone network, denoted as... , ,in, Represents the feature layer; The number of channels representing the features of the original underwater image; Indicates the height of features in the original underwater image; Indicates the width of the features in the original underwater image; S22. Extract and enhance the features of the image using a backbone network, denoted as... , ,in, The number of channels representing the enhanced image features; Indicates the height of the enhanced image features; This indicates the width of the enhanced image features.

[0052] In a specific implementation, as a preferred embodiment of the present invention, step S3 includes: S31. Extracting underwater raw image features from the backbone network and enhance image features The underwater image serves as the input image for the dual-path feature cross-enhancement module. In this embodiment, since the enhanced underwater image provides appearance and discrimination information beneficial for target detection, while the original underwater image retains the natural features of the scene, effectively utilizing the features of both the enhanced and original images is crucial for underwater target detection. To cross-enhance the features of both, a dual-path feature cross-enhancement module is designed, such as... Figure 4 As shown.

[0053] S32. Reshape the input image using features. The reshaped features are as follows: and The input image is divided into non-overlapping segments. A partial window, in which and All represent the total number of windows. This indicates the window size, which is set to the default value in this embodiment. ; S33. Introduce a feature cross-enhancement mechanism, the formula of which is as follows:

[0054]

[0055] The query for the original image features is multiplied by the key matrix of the enhanced image features, and then a relative positional bias is added. The attention score is used to dynamically adjust the attention score, thereby reflecting the relative positional relationship between elements in the input sequence; then, softmax is used to normalize the result, generating the attention score, which is then multiplied by the value matrices of the original image features and the enhanced image features, respectively; in the above formula, Represents the query, key, and value mapped by features; Features of the original underwater image Queries, keys, and values; Enhance image features respectively Queries, keys, and values; S34, calculated WCA Features of the original image and features of the enhanced image The output features of the layer are as follows:

[0056] in, and They represent Features of the original image and features of the enhanced image The output features of the layer; Indicates normalization; S35, in WCA Based on this, the residual block output features of the original image features and the MLP module are calculated. The layer output features are as follows:

[0057]

[0058] in, and The residual block output features representing the original image features and the MLP module's features are respectively... Layer output features; S36, In WCA Based on this, the residual block output features of the enhanced image features and the MLP module are calculated. The layer output features are as follows:

[0059]

[0060] in, and These represent the residual block output features for enhancing image features and the MLP module, respectively. Layer output features.

[0061] In this embodiment, for features derived from enhanced images It works by comparing the original image with... and Cross-weighted dot product is used to fuse the information from both images, while residual connections preserve information from the enhanced image and the original image features. Similarly, it works by interacting with images from the enhanced image. and Cross-weighted dot product is used to fuse the information from both images, while residual connections preserve the information from the original image.

[0062] In a specific implementation, as a preferred embodiment of the present invention, step S4 includes: S41. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the width dimension, as follows:

[0063] in, This represents the feature layer to be encoded in the dual-branch 3D adaptive feature fusion module. In the In terms of width The value at that location, This represents the feature layer obtained after encoding along the width dimension. exist The value at that location, These represent the batch number, channel number, height, and width of the feature layer, respectively. Represents the weight values ​​during the encoding process; S42. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the height dimension, as follows:

[0064] in, This represents the feature layer obtained after encoding along the height dimension. exist The value at that location,

[0065] S43. The adaptive weights encoded along the width and height dimensions are calculated using the activation function, as shown in the following formula:

[0066] in, Representation of features and Adaptive weights in the height dimension; Representation of features and Adaptive weighting in the width dimension; Representing dimensions and dimensions Exchange operations between them; Representing dimensions and dimensions Exchange operations between them; Representation of features and ; S44. Copy the adaptive weights encoded along the width dimension along the width direction, and copy the adaptive weights encoded along the height dimension along the height direction, so that their shape becomes... ; S45. The fused features are calculated using the weights encoded along the width and height dimensions, respectively. The calculation formula is as follows:

[0067] in, and They represent and The fused features are obtained by fusing along the height and width dimensions; Represents element-wise multiplication. , These are features from the original image and the enhanced image, respectively; S46. Calculate the fusion features using the activation function. and The adaptive weights encoded along the width and height dimensions are formulated as follows:

[0068]

[0069] S47, Based on fusion features and Adaptive weights encoded along the width and height dimensions, adaptive fusion features and The final output features are calculated as follows:

[0070] in, The final output features after adaptive fusion .

[0071] In this embodiment, the image enhancement result may contain noise contamination, potentially leading to poor detection performance after fusing the original and enhanced image features. A dual-branch 3D adaptive feature fusion module is designed, such as... Figure 5 As shown, information across the channel, height, and width dimensions of the features is encoded separately to generate adaptive weights. By controlling the fusion of the two branch features through weighting and selective manipulation, the target features are selectively focused on while ignoring noise, reducing noise interference during feature fusion after image enhancement, and finally fusing the features of the two branches after feature cross-enhancement together.

[0072] In specific implementation, as a preferred embodiment of the present invention, such as Figure 6 As shown, the method of the present invention further includes: The steps for training the model using UnitModule unsupervised loss and detection loss as the total loss function are as follows:

[0073] in, Indicates detection loss, This indicates unsupervised loss. In this embodiment, the weighting coefficients are represented. It is 0.5.

[0074] The visualization results of the method of this invention are shown in the following figure. Figure 6 As shown in the figure, the left side represents the ground truth labels, the middle side represents the detection results of the baseline model, and the right side represents the detection results of the method of this invention. It can be seen from the figure that the improved model, which incorporates underwater image enhancement, has significantly fewer false positives compared to the original model, thus improving detection accuracy.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target detection method combining underwater image enhancement, characterized in that, include: S1. Enhance the original underwater image using UnitModule to generate an enhanced image with clearer object boundaries; S2. Use a backbone network to extract features from the original underwater image and the enhanced image, respectively; S3. Input the extracted underwater raw image features and enhanced image features into the dual-path feature cross-enhancement module to perform feature cross-enhancement; S4. The features of the two branches after cross-enhancement are fused by the dual-branch 3D adaptive feature fusion module to filter out noise generated during image enhancement; S5. After passing through the neck network, the detection head outputs the detection results.

2. The target detection method combining underwater image enhancement according to claim 1, characterized in that, The UnitModule enhances images based on the inverse process of the improved Koschmieder model, specifically as follows: The original underwater image is input into the UnitModule as a degraded image to learn the transport map and the underwater global background light; Based on the learned transmission map and underwater global background light, the enhanced image is calculated using the following formula: in, This indicates an undegraded image, i.e., an enhanced image; This indicates a degraded image, i.e., the original underwater image; This represents the transmission diagram, which physically indicates the proportion of light that reaches the visual sensor after particle attenuation. This represents the underwater global background light, which is the average value of each channel in the input image.

3. The target detection method combining underwater image enhancement according to claim 2, characterized in that, The unsupervised learning strategy of the UnitModule is as follows: Input a degraded image, denoted as For degraded images Set a hyperparameter The image was further degraded by calculation. The calculation formula is as follows: By minimizing and The difference between them yields the transport graph loss, as follows: in, Represents the unsupervised loss of the transport graph; Represents pixels; Indicates based on degraded image The estimated value of UnitModule; Indicates based on degraded image Further degraded images The estimated value of UnitModule; Indicates control The hyperparameters for the degree of degradation range from 0 to 1. The key to achieving transport graph loss is to pay attention to the image... and images Underwater global background light Exactly the same.

4. The target detection method combining underwater image enhancement according to claim 1, characterized in that, Step S2 includes: S21. Extract features from the original underwater image using a backbone network, denoted as... , ,in, Represents the feature layer; The number of channels representing the features of the original underwater image; Indicates the height of features in the original underwater image; Indicates the width of the features in the original underwater image; S22. Extract and enhance the features of the image using a backbone network, denoted as... , ,in, The number of channels representing the enhanced image features; Indicates the height of the enhanced image features; This indicates the width of the enhanced image features.

5. The target detection method combining underwater image enhancement according to claim 1, characterized in that, Step S3 includes: S31. Extracting underwater raw image features from the backbone network and enhance image features The input image serves as the dual-path feature cross-enhancement module; S32. Reshape the input image using features. The reshaped features are as follows: and The input image is divided into non-overlapping segments. A partial window, in which and All represent the total number of windows. Indicates window size; S33. Introduce a feature cross-enhancement mechanism, the formula of which is as follows: The query for the original image features is multiplied by the key matrix of the enhanced image features, and then a relative positional bias is added. The attention score is used to dynamically adjust the attention score, thereby reflecting the relative positional relationship between elements in the input sequence; then, softmax is used to normalize the result, generating the attention score, which is then multiplied by the value matrices of the original image features and the enhanced image features, respectively; in the above formula, Represents the query, key, and value mapped by features; Features of the original underwater image Queries, keys, and values; Enhance image features respectively Queries, keys, and values; S34, calculated WCA Features of the original image and features of the enhanced image The output features of the layer are as follows: in, and They represent Features of the original image and features of the enhanced image The output features of the layer; Indicates normalization; S35, in WCA Based on this, the residual block output features of the original image features and the MLP module are calculated. The layer output features are as follows: in, and The residual block output features representing the original image features and the MLP module's features are respectively... Layer output features; S36, In WCA Based on this, the residual block output features of the enhanced image features and the MLP module are calculated. The layer output features are as follows: in, and These represent the residual block output features for enhancing image features and the MLP module, respectively. Layer output features.

6. The target detection method combining underwater image enhancement according to claim 1, characterized in that, Step S4 includes: S41. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the width dimension, as follows: in, This represents the feature layer to be encoded in the dual-branch 3D adaptive feature fusion module. In the In terms of width The value at that location, This represents the feature layer obtained after encoding along the width dimension. exist The value at that location, These represent the batch number, channel number, height, and width of the feature layer, respectively. Represents the weight values ​​during the encoding process; S42. The dual-branch 3D adaptive feature fusion module uses convolutional layers to encode features along the height dimension, as follows: in, This represents the feature layer obtained after encoding along the height dimension. exist The value at that location, S43. The adaptive weights encoded along the width and height dimensions are calculated using the activation function, as shown in the following formula: in, Representation of features and Adaptive weights in the height dimension; Representation of features and Adaptive weighting in the width dimension; Representing dimensions and dimensions Exchange operations between them; Representing dimensions and dimensions Exchange operations between them; Representation of features and ; S44. Copy the adaptive weights encoded along the width dimension along the width direction, and copy the adaptive weights encoded along the height dimension along the height direction, so that their shape becomes... ; S45. The fused features are calculated using the weights encoded along the width and height dimensions, respectively. The calculation formula is as follows: in, and They represent and The fused features are obtained by fusing along the height and width dimensions; Represents element-wise multiplication. , These are features from the original image and the enhanced image, respectively; S46. Calculate the fusion features using the activation function. and The adaptive weights encoded along the width and height dimensions are formulated as follows: S47, Based on fusion features and Adaptive weights encoded along the width and height dimensions, adaptive fusion features and The final output features are calculated as follows: in, The final output features after adaptive fusion .

7. The target detection method combining underwater image enhancement according to claim 1, characterized in that, Also includes: The steps for training the model using UnitModule unsupervised loss and detection loss as the total loss function are as follows: in, Indicates detection loss, This indicates unsupervised loss. This represents the weighting coefficient.

Images