Dense multiplexing and jump connection underwater image enhancement method for multilayer color features

Abstract

The invention relates to the technical field of underwater image enhancement, in particular to a dense multiplexing and jump connection underwater image enhancement method for multilayer color features, which comprises the following steps: S1, a feature extraction branch network CFCNN receives an original image, and transmits the original image to each channel splicing layer through a jump connection strategy, extracting features through a densely connected convolution module DCBlock, and extracting advanced features of each color space through a feature extraction branch network CFCNN; and S2, based on an underwater image enhancement model MCF-Net of multilayer color features, features are extracted in three color spaces of RGB, HSV and Lab through a parallel feature extraction branch network CFCNN, an attention mechanism module CBAM is introduced for adaptive feature selection and weighting after channel splicing, image enhancement is realized through convolution operation, and a result is output. The diversity of feature representation is enriched, and the sensitivity of the model to underwater image color features is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater image enhancement, and in particular to an underwater image enhancement method based on dense multiplexing and jump connection of multi-layer color features. Background Art

[0002] Due to the complexity of underwater environments and lighting conditions, underwater images are affected by wavelength-dependent absorption and scattering, often exhibiting degradation issues such as color distortion and low contrast. These phenomena significantly limit the practical application of underwater images in marine biology, archaeology, marine ecology, and other fields. In recent years, underwater image enhancement has been widely studied to improve the visual quality of underwater images with varying degrees of color distortion and contrast reduction.

[0003] Compared with traditional underwater image enhancement methods based on algorithm engineering and physical models, underwater image enhancement methods based on deep learning can directly learn the mapping between degraded images and enhanced images from data without the need for manual parameter adjustment and the combination of prior knowledge about the physical processes involved in underwater image formation. They are highly versatile.

[0004] While an underwater image enhancement method published in publication number CN120013839A incorporates multi-color space transformations and a cross-color space Transformer module, it lacks a dense connection strategy, resulting in low feature reuse efficiency. Furthermore, downsampling operations can easily lead to loss of detail and red artifacts. Consequently, current technologies still have room for improvement in multi-color space feature fusion, global consistency enhancement, and adaptive enhancement of degraded regions. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the existing technology and provide an underwater image enhancement method that uses dense multiplexing and skip connections for multi-layer color features. By extracting features from different color spaces and incorporating an attention mechanism to highlight key features within the extracted features, the method enriches the diversity of feature representations and enhances the model's sensitivity to underwater image color features. Furthermore, when extracting features from each color space, this invention employs a dense connection and skip connection strategy, effectively extracting the potential high-level features of each color space while using a relatively small number of parameters. This improves the model's ability to withstand degradation issues inherent in underwater environments, such as color distortion and reduced contrast.

[0006] The present invention is achieved through the following technical solution: an underwater image enhancement method for dense multiplexing and jump connection of multi-layer color features includes the following steps: S1. The underwater image enhancement model MCF-Net receives the original image and obtains three color space images: RGB, HSV, and Lab. The underwater image enhancement model MCF-Net includes three parallel feature extraction branch networks CFCNN. The three color space images are input into a feature extraction branch network CFCNN respectively. S2, a single feature extraction branch network CFCNN receives the image in the corresponding color space, transmits the image in the corresponding color space to each channel splicing layer through the skip connection strategy, and extracts features through the densely connected convolution module DCBlock; S3. The features extracted by each feature extraction branch network CFCNN are introduced into the attention mechanism module CBAM after channel splicing for adaptive feature selection and weighting, and then image enhancement is achieved through convolution operation to output the results.

[0007] The S1 includes the following sub-steps: S1-1, receiving the original image and performing normalization processing; S1-2, obtaining an HSV color space image and a Lab color space image based on the RGB color space; S1-3. Input images in different color spaces into the corresponding feature extraction branch network CFCNN respectively.

[0008] The original image received in S1-1 is cropped to 620×460 pixels, and the number of channels of the original image is 3.

[0009] The S2 includes the following sub-steps: S2-1. In the feature extraction branch network CFCNN, the image in the corresponding color space is connected to the convolution module DCBlock, and the image in the corresponding color space is directly transmitted to each channel splicing layer through the skip connection strategy; S2-2: The image in the corresponding color space passes through three convolution modules DCBlock. The output of the convolution module DCBlock is the channel splicing of the results after the activation function output of each convolution layer; S2-3. Map the number of feature map channels to 32 as the final output of the feature extraction branch network CFCNN.

[0010] In S2-2, a single convolution module DCBlock convolves the image with 7x7, 5x5, and 3x3 convolution kernels in sequence. After each convolution operation, nonlinearity is introduced through the Leaky ReLU activation function. The output of the convolution module DCBlock is the channel concatenation of the output results of the activation functions of each convolution layer.

[0011] The image in the corresponding color space in S2-3 is spliced through three convolution modules DCBlock and the corresponding channels, and then the number of feature map channels is mapped to 32 through a convolution layer with a 1 x 1 convolution kernel, which serves as the final output of the feature extraction branch network CFCNN.

[0012] The S3 includes the following sub-steps: S3-1, integrating multi-dimensional feature information mined from different color spaces through the channel splicing layer; S3-2, by introducing the attention mechanism module CBAM, adaptive feature selection and weighting are achieved; S3-3. Use continuous convolution operations to further fuse the attention-weighted features, map the number of feature map channels back to 3, and output the enhanced underwater image.

[0013] The attention mechanism module CBAM includes a channel attention CAM and a spatial attention SAM connected in series, and S3-2 includes the following sub-steps: S3-2-1. Channel Attention CAM uses global average pooling and global maximum pooling to capture the global context information of the feature map, and inputs this information into a shared multi-layer perceptron to learn the importance weights of each channel. Channel Attention CAM can assist the MCF-Net model of the present invention in dynamically identifying and enhancing the feature channels that are most sensitive to underwater scenes and contribute the most significantly. S3-2-2. Spatial Attention SAM receives the channel-weighted feature map output by the channel attention CAM, aggregates the channel information by performing average pooling and maximum pooling in the channel dimension, and further generates a spatial attention map to assist MCF-Net in focusing on the most degraded areas in the underwater image.

[0014] The S3-3 includes the following sub-steps: S3-3-1. Further feature fusion is performed through 3x3 convolution kernels to promote more sufficient interaction and nonlinear combination of the extracted key feature information in the spatial dimension, enabling MCF-Net to learn more complex local patterns and contextual relationships. S3-3-2, smoothly integrating features under different attention weights to ensure that the feature representation is highly refined and coordinated before being passed to the final output layer; S3-3-3 uses a 1x1 convolution kernel to reduce the feature map channel dimension, mapping the channel dimension back to 3. A sigmoid activation function is used to ensure that the pixel values of the output image are between [0, 1]. The final two convolutional layers work together to ensure the generation of a high-quality, detailed, and naturally colored enhanced image.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a feature extraction branch network CFCNN based on dense multiplexing and skip connections. By adopting a dense connection strategy, it realizes the direct reuse of deep and shallow network features, significantly enhances the transmission efficiency of information flow, thereby effectively alleviating problems such as gradient disappearance, and ensures that the feature extraction branch network CFCNN fully captures the complex textures and details of underwater images in various color spaces.

[0016] The feature extraction branch network (CFCNN) also introduces a cross-layer skip connection strategy to fuse feature maps at different depths. This promotes the effective integration of global contextual information and local detail features, facilitating the restoration of degradation issues such as contrast reduction and color distortion caused by uneven illumination and scattering in underwater environments. Through this unique synergy of dense multiplexing and multi-level skip connections, the feature extraction branch network (CFCNN) is able to adaptively learn and extract multi-layer features across the color space of underwater images and efficiently transmit them to the subsequent image reconstruction module.

[0017] This paper proposes an underwater image enhancement model, MCF-Net, based on multi-layer color features. By constructing an end-to-end enhancement network that deeply integrates multi-scale color priors with an attention mechanism, the model accurately restores the color and detail information of underwater images. Using three parallel feature extraction branches, the CFCNN, the MCF-Net extracts color information at different levels in the RGB, HSV, and Lab color spaces of underwater images, laying the foundation for subsequent refined enhancement.

[0018] By introducing a dual attention mechanism (CBAM) module, dynamic weights are assigned to different color channels and spatial regions, enabling the underwater image enhancement model MCF-Net to adaptively focus on color components and regions most severely affected by the underwater environment, thereby achieving targeted color restoration and contrast enhancement. Through a carefully designed feature fusion strategy, the underwater image enhancement model MCF-Net effectively integrates multi-layer color features with attention-weighted features, ensuring a synergistic improvement in global color consistency and local detail clarity, effectively overcoming color distortion and detail blurring caused by water absorption and scattering.

[0019] In the overall design process of the underwater image enhancement model MCF-Net, the present invention adopts convolution kernels of various sizes, including 7x7, 5x5, 3x3 and 1x1. This multi-size convolution kernel synergistic design enables the underwater image enhancement model MCF-Net to simultaneously capture multi-level features from large-scale contextual information to local fine textures, effectively addressing various complex degradation problems that may occur in underwater images.

[0020] The underwater image enhancement model, MCF-Net, avoids downsampling throughout the feature extraction and enhancement process. Instead, it uses padding to ensure that the feature map size remains constant throughout the entire process. This design maximizes the preservation of the image's original spatial information and high-frequency details, fundamentally avoiding the red artifacts introduced by feature map size changes in underwater images. Through this strategy, the underwater image enhancement model, MCF-Net, improves feature representation capabilities while maintaining the structural fidelity and natural color of the enhanced results, achieving precise and efficient enhancement of underwater images. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is an example image of the UIEB image dataset of real underwater image data; Figure 2 This is a schematic diagram of the feature extraction branch network CFCNN architecture; Figure 3 This is a schematic diagram of the DCBlock architecture of the convolution module; Figure 4 This is a comparison chart of the activation function ReLU and Leaky ReLU; Figure 5 This is a schematic diagram of the MCF-Net network architecture of the underwater image enhancement model; Figure 6 This is a schematic diagram of the attention mechanism module CBAM architecture; Figure 7 It is a trend chart of the model trained using the training module; Figure 8 It is a qualitative comparison of synthetic underwater image test dataset; Figure 9 It is a qualitative comparison of the real underwater image test dataset; Figure 10 This is an example image of the synthetic underwater image dataset. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0023] Example 1 Reference Figures 1-10 , an underwater image enhancement system with dense multiplexing and skip connections for multi-layer color features, including: Dataset, used to store underwater image information; Three parallel feature extraction branch networks (CFCNNs) process images in RGB, HSV, and Lab color spaces, respectively. Each branch extracts features through densely connected DCBlock modules and transfers the original image to each channel concatenation layer through skip connections, ultimately outputting a feature map. The underwater image enhancement model MCF-Net is used to integrate the features extracted by each branch of the feature extraction branch network CFCNN, introduce the attention mechanism module CBAM for adaptive feature selection and weighting, and output the enhanced image after feature fusion and channel dimensionality reduction; Training module, using MSE loss L MSE , structural similarity loss L SSIM and VGG perceptual loss L VGG The weighted combination of is used as the total loss function to train the underwater image enhancement model MCF-Net.

[0024] Wherein, the image dataset includes a real underwater image dataset UIEB and a synthetic underwater image dataset; The real underwater image dataset UIEB contains multiple real underwater images taken under different lighting conditions; The synthetic underwater image dataset utilizes an underwater degradation physical model to perform degradation processing on clear land images, synthesizes underwater images, and generates a dataset; The image dataset stores images with a size of 620 × 460 pixels.

[0025] Limited by its data-driven nature, neural network models rely on a large amount of underwater "degraded image-enhanced image" data for training. However, real underwater image datasets with reference enhanced images are relatively rare. Therefore, the present invention combines the real underwater image dataset UIEB with a synthetic underwater image dataset for subsequent training and testing tasks.

[0026] Specifically, the real underwater image dataset UIEB in this embodiment consists of 890 real underwater images taken under different lighting conditions, with different color ranges and contrasts. The synthetic underwater image dataset uses an underwater degradation physical model to degrade clear land images so that the synthetic underwater images present the visual effects of different underwater conditions. Figure 1 and Figure 10 , showing some example images of two underwater image datasets. Figure 1 and Figure 10 The Ⅰ rows in the middle are the original underwater images of the corresponding datasets, and the Ⅱ rows are the reference enhanced image labels provided by the corresponding datasets.

[0027] As can be seen, both datasets cover various degradation issues that may occur in underwater environments, such as color distortion and contrast reduction. To balance model training speed and computer hardware performance, the collected underwater images were cropped to 620×460 pixels.

[0028] The feature extraction branch network CFCNN in this embodiment is used to extract potential high-level features from each color space. Its network architecture is as follows: Figure 2 As shown in Figure 2, the feature extraction branch network (CFCNN) consists of a set of densely connected convolutional modules (DCBlocks). The original input image, a 620×460 pixel underwater image, is transmitted to each channel concatenation layer using skip connections. After convolution with a 3x3 convolutional kernel, the number of feature map channels is converted from 3 to 16. After an activation function, the layers are connected to three DCBlocks. The final convolutional layer uses a 1x1 convolutional kernel to map the feature map channels to 32 for subsequent image reconstruction.

[0029] Specifically, Figure 2 In the above figure, Input represents input, and the feature extraction branch network CFCNN includes multiple convolution modules DCBlock. A single convolution module DCBlock includes a convolution layer-activation function layer with convolution kernels of different sizes. Each convolution module DCBlock is connected through a channel splicing layer, i.e. Concat, which contains the output of the current convolution module DCBlock and the output result of the previous channel splicing layer; The last convolution module DCBlock is connected to a convolution layer with a 1 x 1 convolution kernel through a channel splicing layer, namely CONV-1.

[0030] Reference Figure 3 , its blue blocks represent convolutional layers. A single convolution module DCBlock includes convolutional layers with three sizes of convolution kernels. CONV-7, CONV-5, and CONV-3 represent convolutional layers with convolution kernel sizes of 7x7, 5x5, and 3x3, respectively, to achieve progressive receptive field expansion, thereby being able to capture the inherent multi-scale degradation features in underwater images in a hierarchical manner.

[0031] Each convolutional layer is connected to an activation function layer, which uses a Leaky ReLU activation function, i.e., LReLU; The output of the convolution module DCBlock is the channel splicing of the results after the activation function output of each convolution layer.

[0032] Compared to the widely used ReLU activation function, the Leaky ReLU activation function can effectively alleviate the problem of neuron death. The ReLU activation function sets all negative input values to zero, which causes some neurons to be permanently inactivated during training and no longer respond to any input, resulting in information loss and impeded gradient propagation.

[0033] The Leaky ReLU activation function assigns a small, non-zero positive slope to negative input values. LeakyReLU ensures that the neuron can still generate gradients when the input is negative, ensuring that the model has a more stable gradient flow. The comparison diagram of the two activation functions is as follows Figure 4 shown.

[0034] This embodiment incorporates the idea of dense connection strategy into the design of the entire feature extraction branch network CFCNN, which is expressed as follows: Figure 2 and Figure 3 The brown lines in the figure. For the convolutional module DCBlock that constitutes the feature extraction branch network CFCNN, the dense connection strategy is manifested as the output of DCBlock being the channel splicing of the output results of each convolutional layer after the activation function. This approach not only promotes the deep extraction and efficient reuse of local features, but also strengthens the local information flow, ensuring that the effective details learned within the unit can be transmitted without attenuation, thereby generating a more refined feature representation of the local color degradation pattern. For the macro design of the feature extraction branch network CFCNN, the dense connection is manifested as the channel splicing layer covering not only the output results of the current convolutional module DCBlock, but also the output results of the channel splicing layer of the previous layer. This design achieves the step-by-step enhancement of global contextual features, improves gradient propagation, and enhances the stability and efficiency of network training.

[0035] The feature extraction branch network CFCNN is composed of a large number of convolutional layers. In order to avoid the problem of gradient disappearance as much as possible, the present invention adopts a skip connection strategy to continuously transmit the original input image to each channel splicing layer, which is expressed as Figure 2 As shown by the green lines in the figure, these skip connections give greater weight to channels associated with the original input image, ensuring that the model can learn new features from each processed image.

[0036] Reference Figure 5 and Figure 6 , the attention mechanism module CBAM includes a channel attention CAM and a spatial attention SAM connected in series; The attention mechanism module (CBAM) sequentially connects two sets of convolutional layers and activation function layers. One set of convolutional layers and activation function layers consists of a 3x3 convolutional kernel, corresponding to the LReLU activation function. The other set consists of a 1x1 convolutional kernel, corresponding to the Sigmoid activation function.

[0037] The attention mechanism module CBAM used in this paper is designed to address the problems of information redundancy and insufficient recognition after multi-source feature fusion, so that the model can adaptively focus on the key information that enhances the underwater image. Its network architecture is as follows: Figure 6 As shown in Figure 2, CBAM connects channel attention CAM and spatial attention SAM in series to provide stronger feature selection and expression capabilities. The network architecture of the underwater image enhancement model MCF-Net is shown in Figure 2. Figure 5 As shown in the figure, at the end of the underwater image enhancement model MCF-Net, two sets of convolutional layer-activation function layers are used to achieve the final feature integration and image reconstruction.

[0038] In order to train the underwater image enhancement model MCF-Net, this embodiment adopts a weighted combination of three types of loss functions as the total loss function, namely MSE loss , structural similarity loss and VGG perceptual loss , the total loss function The calculation formula is as follows: ; in, , , are the weight coefficients of the three loss functions respectively.

[0039] MSE loss Underwater enhanced image I and reference image I output by the computational model * The sum of the squared differences between the pixels ensures that the enhanced image output by the model retains edges and details. The calculation formula is as follows: ; Structural similarity loss It can effectively measure the perceptual similarity between the underwater enhanced image output by the model and the reference image. The SSIM value is usually calculated on the grayscale image to focus on the brightness, contrast and structural information of the image and reduce the interference of color deviation. Specifically, for each pixel in the image , the corresponding SSIM value is evaluated in a neighboring pixel block of 13×13 pixels, and the calculation formula is as follows: ; in, and The enhanced graph output by the model is The mean and standard deviation of the pixel block centered at . Correspondingly, and The reference image is The mean and standard deviation of the pixel block centered at . is the covariance between the two pixel blocks mentioned above. and To avoid small constants that produce division by zero errors, set =0.02, =0.03. Finally, the SSIM loss The calculation formula is as follows: ; VGG perceptual loss The goal is to evaluate the similarity of images by comparing their representations in deep feature space, so as to better capture the high-level semantic information perceived by the human visual system. Specifically, the VGG-19 network pre-trained on the ImageNet dataset is used as the feature extractor. When calculating the perceptual loss, the enhanced image I output by the model is compared with the reference image I. * They are input into the VGG-19 network respectively, and then the feature maps output by the specific convolutional layer in the network are extracted as the deep feature representations of the two images. The calculation method of VGG perceptual loss is defined as the Euclidean distance between the two feature representations, and its calculation formula is as follows, where is the j-th convolutional layer of the pre-trained network VGG-19.

[0040] .

[0041] In the three different types of loss functions, the subscript i represents the pixel value of the i-th pixel in the image.

[0042] Compared to images of terrestrial scenes, the color deviations of underwater images exhibit a wider range and more complex characteristics, ranging from classic blue and green to less common yellow. This diverse color distortion caused by the selective absorption and scattering of water makes it difficult for network architectures that use a single color space to fully capture and correct all types of degradation.

[0043] Inspired by the idea that traditional image enhancement algorithms often perform targeted processing in different color spaces, the present invention uses three parallel color space feature extraction branch networks (CFCNNs) to perform deep feature mining in three representative color spaces: RGB, HSV, and Lab. Among them, the RGB color space has a clear physical meaning because it directly corresponds to human perception and display devices, which is convenient for image storage and display. However, due to the high correlation between its R, G, and B components, its features are easily affected by the coupling of factors such as illumination changes, occlusion, and shadows, making it difficult to separate pure color information. The HSV color space decouples color information into hue, saturation, and brightness, which can more intuitively reflect the color essence, vividness, and overall brightness of the image, which is very valuable for adjusting the color saturation and contrast of underwater images. The Lab color space has perceptual uniformity, in which the L component represents brightness, and the a and b components represent color opposition dimensions (green-red and blue-yellow). This design makes the color distribution more uniform, can cover all color ranges that the human eye can perceive, and its brightness is relatively independent of color information, which is conducive to color correction regardless of brightness.

[0044] Given the unique advantages of these three color spaces and their ability to represent different aspects of color information, the underwater image enhancement model MCF-Net designed in this paper integrates these complementary properties into a unified deep learning framework, comprehensively capturing rich information related to underwater image degradation across multiple dimensions, including color, hue, saturation, and brightness. Furthermore, given that different color spaces have varying sensitivities to color differences—meaning that a slight color difference between two pixels in one color space can appear significant in another—MCF-Net's multi-color space parallel processing mechanism enables more effective quantification and perception of subtle color deviations in underwater images.

[0045] Specifically, the underwater image enhancement system with dense multiplexing and jump connection for multi-layer color features includes the following steps when performing image augmentation: S1. The underwater image enhancement model MCF-Net receives the original image and obtains three color space images: RGB, HSV, and Lab. The underwater image enhancement model MCF-Net includes three parallel feature extraction branch networks CFCNN. The three color space images are input into a feature extraction branch network CFCNN respectively to obtain implicit high-level color features. The S1 includes the following sub-steps: S1-1, receiving an original image and performing normalization processing; the original image received in S1-1 is cropped to 620×460 pixels, and the number of channels of the original image is 3.

[0046] S1-2, obtaining an HSV color space image and a Lab color space image based on the RGB color space; S1-3. Input images in different color spaces into the corresponding feature extraction branch network CFCNN respectively.

[0047] S2. The single feature extraction branch network CFCNN receives the image of the corresponding color space, transfers the image of the corresponding color space to each channel splicing layer through the skip connection strategy, and extracts features through the densely connected convolution module DCBlock.

[0048] The S2 includes the following sub-steps: S2-1. In the feature extraction branch network CFCNN, the image in the corresponding color space is connected to the convolution module DCBlock, and the image in the corresponding color space is directly transmitted to each channel splicing layer through the skip connection strategy; S2-2: The image in the corresponding color space passes through three convolution modules DCBlock. The output of the convolution module DCBlock is the channel splicing of the results after the activation function output of each convolution layer. In S2-2, a single convolution module DCBlock convolves the image with 7x7, 5x5, and 3x3 convolution kernels in sequence. After each convolution operation, nonlinearity is introduced through the Leaky ReLU activation function. The output of the convolution module DCBlock is the channel splicing of the output results of the activation function of each convolution layer.

[0049] For a single convolutional module DCBlock, the image undergoes convolution operations with kernel sizes of 7x7, 5x5, and 3x3 in sequence, achieving a progressive expansion of the receptive field, thereby hierarchically capturing the inherent multi-scale degradation features in underwater images.

[0050] Furthermore, the initial large 7x7 convolution kernel can capture a wide range of contextual information and global color distribution, which helps the model understand the degradation process of the entire scene; The subsequent 5x5 convolution kernel effectively perceives regional color deviations and mid-scale structural features based on this; Finally, the 3x3 convolution kernel focuses on extracting fine local texture and edge details.

[0051] This embodiment uses the Leaky ReLU activation function to introduce nonlinearity after the convolution operation, ultimately improving the performance of the model.

[0052] S2-3. Map the number of feature map channels to 32, which is used as the final output of the feature extraction branch network CFCNN. The image corresponding to the color space in S2-3 passes through three convolution modules DCBlock and channel concatenation, and then passes through a convolution layer with a 1 x 1 convolution kernel to map the number of feature map channels to 32, which is used as the final output of the feature extraction branch network CFCNN.

[0053] S3. The features extracted by each feature extraction branch network CFCNN are introduced into the attention mechanism module CBAM after channel splicing for adaptive feature selection and weighting, and then image enhancement is achieved through convolution operation to output the results.

[0054] The S3 includes the following sub-steps: S3-1. For the potential features mined by the feature extraction branch network CFCNN, the multi-dimensional feature information mined from different color spaces is integrated through the channel splicing layer; S3-2. For the rich feature set after splicing, an attention mechanism module CBAM is introduced to realize adaptive feature selection and weighting; the attention mechanism module CBAM includes a channel attention CAM and a spatial attention SAM connected in series. S3-2 includes the following sub-steps: S3-2-1. Channel Attention CAM uses global average pooling and global maximum pooling to capture the global context information of the feature map, and inputs this information into a shared multi-layer perceptron to learn the importance weights of each channel. Channel Attention CAM can assist the MCF-Net model of the present invention in dynamically identifying and enhancing those feature channels that are most sensitive to underwater scenes (such as specific color casts, saturation loss, or uneven brightness) and contribute the most significantly. S3-2-2. Spatial Attention SAM receives the channel-weighted feature map output by Channel Attention CAM, aggregates the channel information by performing average pooling and maximum pooling in the channel dimension, and further generates a spatial attention map to assist the underwater image enhancement model MCF-Net to focus on the most degraded areas in the underwater image.

[0055] S3-3: Use continuous convolution operations to further fuse the attention-weighted features, map the number of feature map channels back to 3, and output the enhanced underwater image. S3-3 includes the following sub-steps: S3-3-1. For the feature maps refined from the dual attention mechanism module CBAM, further feature fusion is performed through a 3x3 convolution kernel to promote more sufficient interaction and nonlinear combination of the extracted key feature information in the spatial dimension, allowing MCF-Net to learn more complex local patterns and contextual relationships. S3-3-2, smoothly integrating features under different attention weights to ensure that the feature representation is highly refined and coordinated before being passed to the final output layer; S3-3-3 uses a 1x1 convolution kernel to reduce the feature map channel dimension, mapping the channel dimension back to 3. A sigmoid activation function is used to ensure that the pixel values of the output image are between [0, 1]. The final two convolutional layers work together to ensure the generation of a high-quality, detailed, and naturally colored enhanced image.

[0056] Example 2 Based on Example 1, in order to train the underwater image enhancement model MCF-Net, this embodiment randomly selects 800 pairs of underwater images from the real underwater image dataset UIEB. The real underwater image dataset UIEB includes a total of 890 real underwater images and corresponding reference images. In addition, the present invention also integrates 1250 synthetic underwater images from the synthetic underwater image dataset, which covers 10 water body types. Therefore, a total of 2050 "underwater image-reference image" pairs are selected as the training dataset of the underwater image enhancement model MCF-Net, and 990 underwater images with different water characteristics are selected from the two datasets as test data. In order to take into account both hardware performance and model efficiency, the images in the training dataset are cropped to a size of 620×460.

[0057] This embodiment implements the training and testing of the underwater image enhancement model MCF-Net on the PyTorch platform. During the model training phase, the Adam optimizer is selected for parameter optimization, and its initial learning rate is set to 0.0001. The two momentum parameters of the Adam optimizer are set to 0.9 and 0.999, respectively. The learning rate adjustment strategy adopts the cosine annealing adjustment strategy, in which the maximum learning rate is 0.1, the minimum learning rate is 0.0001, and the learning rate adjustment cycle is 200. For each loss function weight coefficient, it is set to 1. In addition, in the complete training process, the total number of network iteration training is set to 100. In order to ensure the traceability of the training process, the current training parameters of the model will be saved once every 5 rounds of training. In addition, in order to monitor the learning progress and enhancement effect of the model in real time, the images in the test set will be visualized to display their enhancement results every 500 iterations during the training process.

[0058] Figure 5This article illustrates the overall network architecture of the underwater image enhancement model MCF-Net proposed in this example. The feature extraction branch network CFCNN, which extracts latent feature information from each color space, plays a key role in the entire underwater image enhancement model MCF-Net. To visualize the transformation of the input image throughout MCF-Net, Tables 1 and 2 provide the detailed structural parameters of the feature extraction branch network CFCNN and the complete underwater image enhancement model MCF-Net, respectively.

[0059] Table 1. CFCNN structural parameters of feature extraction branch network

[0060] Table 2 Structural parameters of the underwater image enhancement model MCF-Net

[0061] The evaluation indicators of image enhancement are usually based on the natural statistical characteristics of the image. In order to quantitatively evaluate the performance of the underwater image enhancement model MCF-Net proposed in the present invention, this embodiment adopts two full-reference evaluation indicators and three no-reference evaluation indicators. Among them, in terms of full-reference evaluation indicators, this embodiment adopts peak signal-to-noise ratio PSMR and structural similarity SSIM. In terms of no-reference evaluation indicators, this embodiment adopts underwater image quality metric UIQM, underwater color image quality evaluation indicator UCIQE and information entropy Entropy. In addition, in order to comprehensively evaluate the underwater image enhancement model MCF-Net proposed in the present invention, the present invention compares the proposed method with CLAHE, DCP, GC, ICM, UCM, UDCP and ULAP algorithms. The underwater image enhancement methods used for comparison include image enhancement methods based on traditional algorithm engineering and physical models, which are representative in the field of underwater image enhancement.

[0062] Figure 8 and Figure 9 The enhancement effects of the present application and seven comparison methods on a synthetic underwater test dataset and a real underwater test dataset are respectively demonstrated. (a) and (b) represent the original underwater image and the reference enhanced image labels provided by the dataset, respectively. (c), (d), (e), (f), (g), (h), and (i) represent the enhancement effects of using the CLAHE, DCP, GC, ICM, UCM, UDCP, and ULAP algorithms, respectively. (j) is the enhancement effect of using the present application. The synthetic underwater image dataset uses an underwater degradation physical model to achieve degradation processing of high-quality land images. Its advantage is that targeted underwater degraded images can be obtained by adjusting the corresponding parameters. Figure 8The three test images shown represent the common degradation problems in underwater environments, such as reduced brightness, blue bias, and green bias. However, for these compound degradation problems, it is difficult for the comparison method to solve them at the same time. Algorithms such as CLAHE, GC, and ICM significantly improve the image brightness and solve the problem of reduced brightness, but are helpless against the color distortion problem. Due to the limitations of their own usage scenarios, algorithms such as DCP and UDCP even have a counterproductive effect, resulting in worse visualization of the enhanced results. The underwater image enhancement model MCF-Net proposed in the present invention can solve the compound degradation problem in the synthetic image dataset. It can be seen that the enhanced image eliminates the color distortion problem while retaining its original color characteristics, and also makes a good supplement to the image brightness, achieving the best comprehensive visual effect. Compared with the synthetic underwater image dataset, Figure 9 The presented dataset of real underwater images reflects the true state of underwater environments. It can be seen that all three images tested exhibit varying degrees of color distortion. However, the overall performance of the comparison methods remains poor. For example, algorithms such as UCM and ULAP overcompensate, resulting in red artifacts in the enhanced images. In contrast, the proposed method performs well, with excellent visualization of the enhanced underwater images and a high degree of similarity to the reference enhanced image labels.

[0063] To further validate the performance of the proposed underwater image enhancement model, MCF-Net, we quantitatively evaluated 990 images in the test set using the five evaluation metrics described above. Furthermore, to further demonstrate the versatility of the enhancement method and mitigate the impact of extreme data, we removed the 10 best and 10 worst test data sets and averaged the remaining results. The test results are shown in Table 3. The upward arrow to the right of the evaluation metric indicates that the evaluation result is proportional to the numerical value; that is, the larger the numerical value, the better the enhancement effect. Table 3 Comparison results of objective indicators of various methods on the test dataset

[0064] Table 3 shows the objective metric comparison results of the method of the present invention and the seven comparison algorithms described above on a test dataset. PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) serve as reference evaluation metrics, requiring comparison with the reference images provided in the test dataset to obtain evaluation scores. PSNR scores images by comparing the error and color differences between them, while SSIM scores images by comparing brightness, contrast, and overall structure. The data in the table shows that the method of the present invention achieves the highest scores in both PSNR and SSIM, indicating that the enhanced images of the present invention are close to the reference images in terms of color distribution, brightness distribution, and contrast distribution, fully demonstrating the high versatility of the underwater image enhancement model MCF-Net for underwater image enhancement.

[0065] In addition, Entropy, UIQM (Underwater Image Quality Metric) and UCIQE (Underwater Color Image Quality Evaluation) are used as non-reference evaluation indicators. Mathematical formulas can be used to directly evaluate the quality of the restored underwater images. This also means that the non-reference evaluation indicators can only evaluate some features of the enhanced images. Among them, Entropy is used to measure the amount of information contained in the image, UIQM is used to measure the contrast, color and clarity of the image, and UCIQE is used to measure the chroma, saturation and contrast of the image. According to the data in the table, it can be seen that the MCF-Net proposed in the present invention achieved the highest scores in the two indicators of UCIQE and Entropy, and performed poorly on UIQM. However, this result is different from Figure 8 , Figure 9 The significant disparity in the displayed visualization results demonstrates that relying solely on a single reference evaluation metric is insufficient to directly assess the quality of an enhanced image. A comprehensive evaluation of various evaluation metrics and subjective assessments is crucial for a holistic assessment of the image quality. Therefore, this comprehensive evaluation demonstrates that the present invention effectively addresses inherent issues with underwater images, demonstrating excellent enhancement effectiveness and versatility.

[0066] The above descriptions are merely optional embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention specification under the concept of the present invention, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present invention.

Claims

1. An underwater image enhancement method based on dense multiplexing and skip connections of multi-layer color features, characterized by: The steps include: S1. The underwater image enhancement model MCF-Net receives the original image and obtains three color space images: RGB, HSV, and Lab. The underwater image enhancement model MCF-Net includes three parallel feature extraction branch networks CFCNN. The three color space images are input into a feature extraction branch network CFCNN respectively. S2, a single feature extraction branch network CFCNN receives the image in the corresponding color space, transmits the image in the corresponding color space to each channel splicing layer through the skip connection strategy, and extracts features through the densely connected convolution module DCBlock; S3. The features extracted by each feature extraction branch network CFCNN are introduced into the attention mechanism module CBAM after channel splicing for adaptive feature selection and weighting, and then image enhancement is achieved through convolution operation to output the results.

2. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 1 is characterized in that: The S1 includes the following sub-steps: S1-1, receiving the original image and performing normalization processing; S1-2, obtaining an HSV color space image and a Lab color space image based on the RGB color space; S1-3. Input different color space images into the corresponding feature extraction branch network CFCNN respectively.

3. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 2 is characterized in that: The original image received in S1-1 is cropped to 620×460 pixels.

4. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 2, characterized in that: The S2 includes the following sub-steps: S2-1. In the feature extraction branch network CFCNN, the image in the corresponding color space is connected to the convolution module DCBlock, and the image in the corresponding color space is directly transmitted to each channel splicing layer through the skip connection strategy; S2-2, the image in the corresponding color space passes through three convolution modules DCBlock. The output of the convolution module DCBlock is the channel splicing of the results after the activation function output of each convolution layer; S2-3. Map the number of feature map channels to 32 as the final output of the feature extraction branch network CFCNN.

5. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 4, characterized in that: In S2-2, a single convolution module DCBlock convolves the image with 7x7, 5x5, and 3x3 convolution kernels in sequence. After each convolution operation, nonlinearity is introduced through the Leaky ReLU activation function. The output of the convolution module DCBlock is the channel concatenation of the output results of the activation functions of each convolution layer.

6. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 5, characterized in that: The image in the corresponding color space in S2-3 passes through three convolution modules DCBlock and channel splicing, and then the number of feature map channels is mapped to 32 through a convolution layer with a 1 x 1 convolution kernel, which serves as the final output of the feature extraction branch network CFCNN.

7. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 1, characterized in that: The S3 includes the following sub-steps: S3-1, integrating multi-dimensional feature information mined from different color spaces through the channel splicing layer; S3-2, by introducing the attention mechanism module CBAM, adaptive feature selection and weighting are achieved; S3-3. Use continuous convolution operations to further fuse the attention-weighted features, map the number of feature map channels back to 3, and output the enhanced underwater image.

8. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 7, characterized in that: The attention mechanism module CBAM includes a channel attention CAM and a spatial attention SAM connected in series, and S3-2 includes the following sub-steps: S3-2-1, Channel Attention CAM uses global average pooling and global maximum pooling to capture the global context information of the feature map, and inputs this information into a shared multi-layer perceptron to learn the importance weights of each channel; S3-2-2. Spatial attention SAM receives the channel-weighted feature map output by channel attention CAM, aggregates channel information by performing average pooling and maximum pooling in the channel dimension, and generates a spatial attention map.

9. The underwater image enhancement method based on dense multiplexing and skip connections for multi-layer color features according to claim 7, characterized in that: The S3-3 includes the following sub-steps: S3-3-1. Further feature fusion is performed through a 3x3 convolution kernel to promote more complete interaction and nonlinear combination of the extracted key feature information in the spatial dimension, enabling MCF-Net to learn more complex local patterns and contextual relationships. S3-3-2, smoothly integrating features under different attention weights to ensure that the feature representation is highly refined and coordinated before being passed to the final output layer; S3-3-3. Use a 1x1 convolution kernel to reduce the dimension of the feature map channel, map the channel dimension back to 3, and use the sigmoid activation function to ensure that the pixel value of the output image is between [0, 1].

Images