An underwater salient object detection method based on multi-color space features, a program, a device and a storage medium

Abstract

The present application belongs to the technical field of image processing, and particularly relates to a method, program, device and storage medium for underwater salient target detection based on multi-color space features. 2 The present application introduces YCbCr and HSV color spaces with independent physical meanings, extracts complementary color features from different dimensions, and expands the diversity of color representation. According to the functional differentiation rule of the Net encoder "shallow layer for details, middle layer for structure, and deep layer for semantics", a differentiated layered injection scheme of "shallow layer fusion, middle layer YCbCr, and deep layer HSV" is established, adaptive channel screening is realized through a convolution attention module, dynamic intensity adjustment is realized through a gating fusion module, and a complete access mechanism from channel screening to intensity control is established. The present application can more accurately detect the salient target area in the turbid water environment under artificial lighting, and has higher adaptability to complex underwater environments.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method, program, device, and storage medium for detecting underwater salient targets based on multi-color space features. Background Technology

[0002] When performing missions, AUVs rely primarily on visual imaging for near-range target perception. However, unlike terrestrial imaging environments, underwater optical imaging is affected by both absorption and scattering in water: different wavelengths of light attenuate at significantly different rates, with red light attenuating faster than blue and green light, resulting in widespread color distortion and decreased contrast in images; the scattering effect of suspended particles blurs image boundaries, creating a "fog" effect. Especially in actual AUV operations, more challenging imaging conditions such as artificial lighting and turbid water further exacerbate image degradation, posing entirely new challenges to target detection.

[0003] In the technical approach to object detection, traditional methods focus on category discrimination, relying on predefined category labels to achieve target localization. This is suitable for scenarios where the target category is clear and the sample size is sufficient. However, in actual AUV operations, situations often arise where the target category is unknown and the scene varies greatly. In these cases, salient object detection methods demonstrate greater adaptability. Salient object detection aims to simulate the human visual attention mechanism, automatically extracting the most visually attention-grabbing target region from complex scenes. It can effectively separate the foreground from the background without pre-setting a category, providing foreground priors for subsequent tasks such as recognition, tracking, and grasping.

[0004] Current deep learning-based salient object detection methods (such as...) Methods like BASNet and PoolNet, trained on large-scale land datasets, have achieved good results. However, these methods assume a balanced distribution of RGB three-channel information, allowing for clear distinction between targets and backgrounds through color or texture. When directly applied to underwater scenes, selective absorption by water leads to color information attenuation, and scattering by suspended particles causes boundary blurring, making it difficult for the model to maintain the ability to distinguish targets from the background. This often results in incomplete target region detection, inaccurate boundary localization, and false positives and false negatives. Although there are methods specifically designed for underwater salient target detection, these methods still mainly rely on a single RGB color space for feature extraction, failing to fundamentally solve the problem of color information attenuation caused by color distortion in underwater images. This results in insufficient feature representation capabilities under turbid water and artificial lighting conditions. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of incomplete target detection and inaccurate boundaries caused by the insufficient representation ability of a single RGB color space in underwater environments with color distortion and scattering blur. The invention provides a method, program, device and storage medium for underwater salient target detection based on multi-color space features.

[0006] A method for detecting salient underwater targets based on multi-color space features includes the following steps:

[0007] Input the original underwater RGB image, convert the original underwater RGB image into YCbCr image and HSV image respectively, and extract the high-dimensional color features of YCbCr image and HSV image respectively;

[0008] Through a feature-layered injection strategy, based on The attributes and functions of different levels of the encoder inject the high-dimensional color features of the YCbCr image into... The middle layer of the encoder injects the high-dimensional color features of the HSV image into... The encoder's deeper layers; after concatenating the high-dimensional color features of the YCbCr image and the HSV image along the channel dimension, the data is injected into... The shallow layer of the encoder;

[0009] exist Each injection level of the encoder introduces a convolutional attention module and a gated fusion module. The convolutional attention module adaptively filters the input color features based on channel and spatial dimensions. The gated fusion module dynamically fuses the filtered color features with the original features of the current level of the backbone network to obtain the fused encoded features.

[0010] Will The fused encoded features of the encoder's outputs at each level are input. Decoder The decoder outputs the detection results of underwater salient targets.

[0011] Furthermore, a dual-branch convolutional coding structure is used to extract high-dimensional color features from YCbCr and HSV images;

[0012] Each branch of the dual-branch convolutional coding structure contains two consecutive convolutional blocks, and each convolutional block includes convolution, batch normalization, and ReLU activation functions.

[0013] For YCbCr images, the first convolutional block maps the input YCbCr image to an intermediate feature space and extracts the underlying color structure through nonlinear transformation; the second convolutional block enhances the output of the first convolutional block to a higher dimension, thereby improving semantic expressiveness.

[0014]

[0015]

[0016] in, Image of YCbCr; As an intermediate feature; These are the high-dimensional color features of the YCbCr image.

[0017] Furthermore, the high-dimensional color features of the YCbCr image and the high-dimensional color features of the HSV image are concatenated along the channel dimension and then injected into... The shallow layer of the encoder contains Y components of YCbCr images that carry rich local brightness structure information, and H and S components of HSV images that provide hue and saturation contrast cues. The fusion of the two can enhance the network’s ability to perceive blurred edges under degradation conditions and help the shallow layer capture the general outline of the target even when color information is insufficient.

[0018] The high-dimensional color features of the YCbCr image are injected into... In the middle layer of the encoder, the Y component of the YCbCr image provides the framework of the regional structure, while the Cb and Cr components reflect the color consistency within the region.

[0019] The high-dimensional color features of the HSV image are injected into Deeper into the encoder, the H component of the HSV image provides robust color category information for underwater color bias, enabling the distinction between the foreground and background when their overall colors are similar. The S component of the HSV image reflects the difference in color purity between the foreground and background. Under artificial lighting, the saturation of the foreground decreases due to increased white light, while the saturation of the background increases due to weakened lighting, providing additional criteria for deep semantic discrimination.

[0020] Furthermore, for The encoder's first Original features of layer input If the first If the layer is shallow, then the first layer... Layer output features for:

[0021]

[0022] in, High-dimensional color features of the YCbCr image; High-dimensional color features of HSV images; Indicates concatenation of channel dimensions; This indicates that dynamic fusion is performed through a convolutional attention module and a gated fusion module;

[0023] If the first If the layer is the middle layer, then the first layer... Layer output features for:

[0024]

[0025] If the first If the layer is deep, then the first... Layer output features for:

[0026]

[0027] Furthermore, the adaptive filtering of the input color features by the convolutional attention module in terms of channel and spatial dimensions is specifically as follows:

[0028] The convolutional attention module includes a channel attention module and a spatial attention module;

[0029] The channel attention module is used to distinguish the effectiveness differences of different color component features. The channel attention module focuses on the input color features. Global average pooling and global max pooling are performed to aggregate global information and local most salient responses in the spatial dimension, respectively. The results of these two pooling methods are then fed into a shared multilayer perceptron (MLP) to generate channel weights. ;

[0030]

[0031] in, For activation functions; For global average pooling; This is for global max pooling;

[0032] Channel weight With input color features Element-wise multiplication yields channel enhancement features. ;

[0033]

[0034] in, This represents element-wise multiplication;

[0035] Channel enhancement features output by the channel attention module The spatial attention module performs average pooling and max pooling along the channel dimension, compressing multi-channel features into two single-channel feature maps, reflecting the average and maximum response intensities at each location, respectively. The two feature maps are then concatenated along the channel dimension and convolved with activation functions to generate spatial weights. ;

[0036]

[0037] in, For convolution; Indicates concatenation of channel dimensions;

[0038] Spatial weight With channel enhancement features Element-wise multiplication yields the filtered color features. ;

[0039] .

[0040] Furthermore, the dynamic fusion of the filtered color features with the original features of the current layer of the backbone network via the gating fusion module specifically involves:

[0041] The filtered color features are processed through convolution. Adapt to the original features of the current layer of the backbone network With the same channel dimension, the color features obtained after convolution are... ;

[0042] The original features of the current layer of the backbone network Color features after convolution The weights are concatenated along the channel dimension and then generated through convolution and activation functions to produce gated weights. ;

[0043]

[0044] in, and These are the parameters and biases of the gated convolution, respectively;

[0045] Gating weights For tensors with values ​​between 0 and 1, the dimension is the same as the original features of the current layer of the backbone network. same;

[0046] In the form of residuals, the gating weights are... Weighted color features after convolution Then, the original features of the current layer of the backbone network are superimposed. The fused coding features are obtained. ;

[0047] .

[0048] Furthermore, the aforementioned The decoder uses and The encoder's symmetrical U-shaped structure gradually restores spatial resolution through progressive upsampling and feature stitching.

[0049] Each decoding layer of the decoder first performs bilinear upsampling on the features of the previous layer to make its resolution consistent with the corresponding coding layer. Then, it concatenates the coded features fused with the previous layer along the channel dimension, and then performs feature fusion through convolutional blocks. After progressive upsampling and fusion through multiple decoding layers, the feature resolution is restored to the original input size. Finally, a 1×1 convolution maps the number of feature channels to 1, and the sigmoid activation function constrains the pixel values ​​to the [0,1] interval, outputting a saliency map of the same size as the input image. The closer the pixel value in the saliency map is to 1, the greater the probability that the position belongs to a salient target; the closer it is to 0, the greater the probability that it belongs to the background, thus obtaining the underwater salient target detection result.

[0050] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described underwater salient target detection method based on multi-color space features.

[0051] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described underwater salient target detection method based on multi-color space features.

[0052] A computer program product includes computer instructions that, when executed by a processor, implement the steps of the above-described underwater salient target detection method based on multi-color space features.

[0053] The beneficial effects of this invention are as follows:

[0054] Existing salient target detection methods mostly rely on a single RGB color space. Under underwater conditions of color distortion and uneven illumination, the imbalance of RGB channel information leads to inherent limitations in their color representation capabilities. This invention overcomes this limitation at the feature source level: without adding additional imaging modalities such as depth maps and polarization maps, or sensor hardware, it introduces the YCbCr and HSV color spaces, which have independent physical meaning, to extract complementary color features from different dimensions, thus expanding the diversity of color representation. Meanwhile, existing methods typically employ a uniform injection strategy when introducing external features, ignoring the differentiated requirements of different network layers for feature attributes. Based on the functional differentiation principle of encoders—"shallow layers emphasize details, mid-layers emphasize structure, and deep layers emphasize semantics"—this invention establishes a differentiated layered injection scheme of "shallow fusion, mid-layer YCbCr, and deep-layer HSV." Furthermore, it establishes a complete admission mechanism from channel selection to intensity control through adaptive channel selection via a convolutional attention module and dynamic intensity adjustment via a gated fusion module. This invention can more accurately detect significant target areas in turbid water environments under artificial lighting. The completeness of the detection results and the accuracy of boundary positioning are superior to existing methods, and it has higher adaptability to complex underwater environments. Attached Figure Description

[0055] Figure 1 This is a diagram of the overall network architecture of the present invention.

[0056] Figure 2 This is a schematic diagram of the multi-color space feature extraction module in this invention.

[0057] Figure 3 This is a schematic diagram of the convolutional attention module in this invention.

[0058] Figure 4 This is a schematic diagram of the gating fusion module in this invention.

[0059] Figure 5 Example of underwater degradation image (simulated black box) used in the experiment of this invention.

[0060] Figure 6 This is a comparison chart showing the performance of the present invention with existing underwater salient target detection algorithms.

[0061] Figure 7 This table compares objective evaluation metrics among different algorithms. Detailed Implementation

[0062] The present invention will now be further described with reference to the accompanying drawings.

[0063] This invention provides a method for underwater salient target detection based on multi-color space features. First, a multi-color space feature extraction and selection method is proposed. Based on an RGB color space image, the RGB image is simultaneously converted into YCbCr and HSV images. Multiple features are extracted from the three images. To address the feature redundancy problem caused by multiple features, a feature selection method is designed. Then, target feature injection is proposed. Layered injection strategy for the network. Based on The different attributes and functions of the shallow, middle, and deep layers of the encoder guide different target features to the most suitable layer, thereby improving... The effectiveness of online learning. Finally. The network, based on encoder features, outputs underwater salient target detection results through a decoder to obtain a target saliency map. This invention can more accurately detect salient targets in turbid water environments under artificial lighting. The structural integrity and boundary localization accuracy of the detection results are superior to existing methods, and it has higher adaptability to complex underwater environments.

[0064] Step (1): Multi-color space feature extraction;

[0065] In underwater environments, selective absorption by water leads to a severe imbalance in the RGB three-channel information (red light decays rapidly, while blue and green light dominate). Furthermore, under artificial lighting, brightness and color changes are coupled in the RGB space, making it difficult for models to distinguish color attributes from lighting changes and easily misclassifying areas of enhanced illumination as significant targets. The color representation capabilities of a single RGB space are limited, necessitating the introduction of other color spaces to supplement color information from different perspectives.

[0066] To address the aforementioned problems, this invention designs a Multi-Color Space Feature Extraction Module (MCFEM). By introducing YCbCr and HSV color spaces with different physical meanings, it extracts complementary color features from two dimensions: brightness-chromaticity separation and hue-saturation perception, thereby expanding the color representation sources of the single RGB space. For example... Figure 2 As shown, the input of this module is the original underwater RGB image, and the output is YCbCr color features and HSV color features.

[0067] First, a color space conversion is performed. For the input RGB image, a linear transformation is used to convert it to the YCbCr color space. The YCbCr space decouples luminance and chrominance information. The Y component reflects the light and dark structure, is sensitive to changes in illumination but contains rich structural information; the Cb and Cr components represent the concentration shifts of blue and red, respectively, reflecting color category information and exhibiting good robustness to underwater color shifts. The conversion formula is as follows:

[0068]

[0069] Simultaneously, a non-linear mapping method is used to convert RGB images to the HSV color space. The HSV space independently decomposes color categories, vibrancy, and brightness, which is closer to the way human vision perceives colors. The brightness V is calculated as follows:

[0070]

[0071] Here, Δ is the difference between the maximum and minimum values ​​of the three RGB components, reflecting the magnitude of color difference.

[0072] The saturation S is calculated as follows:

[0073]

[0074] The calculation method for hue H is as follows:

[0075]

[0076] Both of the aforementioned color space conversions are mature technologies, but their combination, when applied to underwater salient target detection tasks, can provide complementary information from different dimensions that the RGB space does not possess.

[0077] Then, a dual-branch convolutional coding structure is used to extract high-dimensional features from the converted YCbCr and HSV images, respectively. Each branch contains two consecutive convolutional blocks, each consisting of a 3×3 convolution, batch normalization, and a ReLU activation function. The first convolutional layer maps the three-channel input to an intermediate feature space, extracting the underlying color structure through nonlinear transformation; the second convolutional layer further enhances the features to a higher dimension, improving semantic expressiveness. Taking the YCbCr branch as an example, this process can be represented as follows:

[0078]

[0079]

[0080] in, It is a 32-dimensional intermediate feature. This is the final output of 64-dimensional color features.

[0081] Similarly, the HSV branch also uses the same two-layer convolutional coding structure to jointly model the H, S, and V channels.

[0082] The two branches are trained independently, sharing the same network structure but learning their parameters independently. This allows the network to automatically learn which channels and dimensions in each color space are most effective for underwater saliency discrimination based on the statistical characteristics of the training data. For example, in scenarios with severe color distortion, the Cb and Cr components of YCbCr may receive higher weights, while in scenarios with uneven lighting, the H component of HSV may play a major role. Through the aforementioned multi-color space feature extraction module, this invention overcomes the representation bottleneck of a single RGB space at the feature source level without increasing sensor hardware or introducing additional imaging modalities.

[0083] Step (2): Layered feature injection;

[0084] The multi-color space feature extraction module provides two complementary high-dimensional color features, YCbCr and HSV. However, effectively integrating this supplementary information into the backbone network is a key factor determining the method's performance. Existing methods typically employ single-location injection or uniform injection across all layers, but neither considers the functional differences between different network layers: shallow layers have high feature resolution and small receptive fields, responsible for details such as edge textures; mid-layers have medium feature resolution, responsible for structural modeling and contour extraction; and deep layers have low feature resolution and large receptive fields, responsible for semantic discrimination. Indiscriminately injecting multi-color space features into all layers leads to a mismatch between feature attributes and layer functions. For example, injecting the brightness structure details of YCbCr into deep layers not only fails to aid semantic abstraction but may also introduce texture noise unrelated to high-level semantics.

[0085] To address the above problems, this invention proposes a hierarchical feature injection strategy, based on... Based on the functional differentiation patterns of the six levels of the encoder, a differentiated allocation scheme is established. The encoder consists of six coding layers (Encoder 1-6). As the number of layers increases, the feature resolution halves with each layer, the receptive field increases progressively, and the semantic abstraction level gradually improves. This invention, based on the functional positioning of each layer, injects color features of different attributes into the most suitable layer:

[0086] (1) Shallow Stage (Encoder 1-2): Shallow features have high resolution and small receptive field, responsible for extracting low-level visual information such as edges and textures, and are sensitive to local details. YCbCr features and HSV features are concatenated along the channel dimension and injected into this stage. The Y component of YCbCr carries rich local brightness structure information, while the H and S components of HSV provide hue and saturation contrast cues. The fusion of the two can enhance the network's ability to perceive blurred edges under degradation conditions, helping the shallow layer to capture the general outline of the target even when color information is insufficient.

[0087] (2) Middle Layer Stage (Encoder 3-4): The feature resolution of the middle layer is lower than that of the shallow layers, and it undertakes the tasks of structural modeling and local contour extraction. At this time, the features have begun to abstract from pixel-level edges to region-level structures, requiring stable regional structure representation and color distribution information. YCbCr features are injected into this stage separately. Its Y component provides the framework of the regional structure, while the Cb and Cr components reflect the color consistency within the region. Compared with the RGB space, the brightness and chromaticity separation characteristics of YCbCr make the middle layer network less susceptible to color distortion interference when modeling regional structures, and it can extract the overall shape of the target more stably.

[0088] (3) Deep Stage (Encoder 5-6): Deep features have the lowest resolution and the largest receptive field, and are responsible for foreground-background semantic discrimination. At this stage, the network needs to move away from low-level edge texture dependence and rely on stable color category information for global semantic feature extraction. HSV features are injected into this stage separately. The H component provides robust color category information for underwater color bias, which can distinguish between the foreground and background when their overall colors are similar. The S component reflects the difference in color purity between the foreground and background. Under artificial lighting, the saturation of the foreground decreases due to increased white light, while the saturation of the background increases due to weakened lighting. This rule provides additional discrimination criteria for deep semantic discrimination.

[0089] Through the aforementioned differentiated allocation strategy, each color feature is directed to the network layer where it can play its most effective role, achieving hierarchical matching between external features and backbone network functions, and structurally avoiding feature mismatch problems caused by uniform injection.

[0090] for The encoder's first Original features of layer input If the first If the layer is shallow, then the first layer... Layer output features for:

[0091]

[0092] in, High-dimensional color features of the YCbCr image; High-dimensional color features of HSV images; Indicates concatenation of channel dimensions; This indicates that dynamic fusion is performed through a convolutional attention module and a gated fusion module;

[0093] If the first If the layer is the middle layer, then the first layer... Layer output features for:

[0094]

[0095] If the first If the layer is deep, then the first... Layer output features for:

[0096]

[0097] Step (3): Adaptive filtering and dynamic fusion;

[0098] The hierarchical injection strategy clarifies which color features to inject into which level, but the quality of the injected features and the fusion method are equally crucial. The effectiveness of different channels in multi-color space features varies significantly underwater: some channels can stably distinguish between the target and the background, while others may carry noise due to scattering interference. If injected directly without screening, noise from ineffective channels will mix with effective information, leading to spurious responses. Furthermore, color features and backbone network features differ in scale and channel dimension, making effective synergy difficult to achieve through direct splicing or fixed-weight fusion.

[0099] To address the aforementioned issues, this invention introduces a Convolutional Attention Module (CBAM) and a Gated Fusion Module (GIM) at each injection level.

[0100] like Figure 3 As shown, for the multi-color space features to be injected CBAM performs channel attention and spatial attention filtering sequentially. The channel attention module distinguishes the effectiveness differences of features from different color components: it performs global average pooling and global max pooling on the input features to aggregate global information and local most salient responses in the spatial dimension, respectively. Then, the results of the two pooling operations are fed into a shared multilayer perceptron (MLP) to generate channel weights. This weight assigns a coefficient between 0 and 1 to each feature channel, suppressing channels that become ineffective due to underwater degradation and enhancing channels that still reflect the difference between the target and the background. The calculation process for channel attention can be expressed as follows:

[0101]

[0102] Channel weight With input color features Element-wise multiplication yields channel enhancement features. ;

[0103]

[0104] Spatial attention module enhances channel attention output features Based on this, it is used to locate key target regions and suppress background interference caused by suspended particles. It performs average pooling and max pooling along the channel dimension, compressing multi-channel features into two single-channel feature maps, reflecting the average response intensity and maximum response intensity at each location, respectively; then, these two feature maps are concatenated along the channel dimension, and spatial weights are generated by 7×7 convolution and sigmoid activation. Spatial weights apply varying degrees of attention to different spatial locations in the feature map, enabling the network to focus on the target region while suppressing background noise. The calculation process for spatial attention can be represented as follows:

[0105]

[0106] Spatial weight With channel enhancement features Element-wise multiplication yields the filtered color features. ;

[0107]

[0108] Through the processing of the aforementioned channel attention and spatial attention modules, the CBAM module achieves adaptive filtering of multi-color space features, making the subsequently injected features more discriminative and stable.

[0109] like Figure 4 As shown, the filtered color features Features of the current layer of the backbone network Dynamic fusion is performed using the GIM module. Since color features and backbone network features differ in numerical distribution and feature scale, direct fusion would lead to optimization difficulties. Therefore, a 3×3 convolution is first used to... Adapt to With the same channel dimension, we get Then, the gating fusion module dynamically determines the injection intensity of the color features based on the joint state of the current backbone network features and color features. Specifically, it will... and The layers are concatenated along the channel dimension, and then gating weights are generated through a 1×1 convolution and a sigmoid function. :

[0110]

[0111] in, Indicates feature splicing, and These are the parameters and biases of the gated convolution, respectively. This is the Sigmoid activation function.

[0112] Gating weights It is a tensor with values ​​between 0 and 1, and its dimension is... Similarly, when color features have a gain effect on the backbone network features of the current level, the weight at the corresponding position approaches 1; when the backbone network features can fully represent the region, or when the color features are strongly affected by degradation, the weight approaches 0. Finally, the gated weighted color features are superimposed on the backbone network features in the form of residuals:

[0113]

[0114] in, This indicates element-wise multiplication.

[0115] The residual structure ensures that the backbone network features always serve as the main representation. Even when the gating weights approach 0, the network can still maintain its original performance, thereby avoiding the color branch from overdoing the network learning and ensuring the training stability during the hierarchical injection process.

[0116] By using adaptive filtering of CBAM and dynamic gating fusion of GIM, this invention ensures that external color features are appropriately integrated into the backbone network and fully utilized to enhance the network's feature representation capabilities.

[0117] Step (4): Output the saliency map;

[0118] The decoder of the backbone network receives the enhanced coding features output from step (3) at each level. The decoder adopts a U-shaped structure symmetrical to the encoder, and gradually restores the spatial resolution through step-by-step upsampling and feature concatenation. Specifically, each decoding layer first performs bilinear upsampling on the features of the previous layer to make its resolution consistent with the corresponding coding layer, and then concatenates it with the enhanced coding features of that level output from step (3) in the channel dimension, and then performs feature fusion through convolutional blocks. After step-by-step upsampling and fusion of 5 decoding layers, the feature resolution is restored to the original input size. Finally, the feature channel number is mapped to 1 by 1×1 convolution, and the pixel value is constrained to the [0,1] interval by the Sigmoid activation function, outputting a saliency map of the same size as the input image. The closer the pixel value in the saliency map is to 1, the greater the probability that the position belongs to a salient target, and the closer it is to 0, the greater the probability that it belongs to the background.

[0119] Example 1:

[0120] To verify the effectiveness of the underwater salient target detection method based on multi-color space features proposed in this invention, the algorithm proposed in this invention is compared experimentally with the BASNet algorithm by Qin et al., the PoolNet algorithm by Liu et al., the F3Net algorithm by Huang et al., and the MINet algorithm by Pang et al. The original underwater images of the black box used in the experiments are attached. Figure 5 As shown in the attached figure, the underwater salient target detection results are compared. Figure 6 As shown.

[0121] Figure 6 (a) shows the truth value label; (b) corresponds to the algorithm of this invention. Figure 5 The result of the processing; (c) the corresponding BASNet algorithm proposed by Qin et al. Figure 5 The result of the processing; (d) corresponding to the PoolNet algorithm proposed by Liu et al. Figure 5 The result of the processing; (e) corresponding to the F³Net algorithm proposed by Huang et al. Figure 5 The result of the processing; (f) corresponding to the MINet algorithm proposed by Pang et al. Figure 5 The result of the processing.

[0122] To objectively evaluate the performance of each method, this invention patent selects four commonly used evaluation indicators for salient target detection: mean absolute error (MAE), maximum F-measure (max F), structural similarity index (S-measure), and enhanced alignment index (E-measure) for quantitative evaluation.

[0123] The evaluation results on the black box pool images are as follows: Figure 7 As shown, a comparative analysis of the evaluation metrics of various algorithms reveals that the algorithm proposed in this invention outperforms other comparative methods in all four metrics: MAE, max F, S-measure, and E-measure. These results demonstrate that the algorithm proposed in this invention exhibits stronger adaptability and higher detection accuracy in underwater environments under artificial lighting conditions in turbid water bodies.

[0124] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting underwater salient targets based on multi-color space features, characterized in that: Input the original underwater RGB image, convert the original underwater RGB image into YCbCr image and HSV image respectively, and extract the high-dimensional color features of YCbCr image and HSV image respectively; Through a feature-layered injection strategy, based on The attributes and functions of different levels of the encoder inject the high-dimensional color features of the YCbCr image into... The middle layer of the encoder injects the high-dimensional color features of the HSV image into... The encoder's deeper layers; after concatenating the high-dimensional color features of the YCbCr image and the HSV image along the channel dimension, it is injected into... The shallow layer of the encoder; exist Each injection layer of the encoder introduces a convolutional attention module and a gated fusion module. The convolutional attention module adaptively filters the input color features based on channel and spatial dimensions. The gated fusion module dynamically fuses the filtered color features with the original features of the current layer of the backbone network to obtain the fused encoded features. Will The fused encoded features of the encoder's outputs at each level are input. Decoder The decoder outputs the detection results of underwater salient targets.

2. The underwater salient target detection method based on multi-color space features according to claim 1, characterized in that: A dual-branch convolutional coding structure is used to extract high-dimensional color features from YCbCr and HSV images; Each branch of the dual-branch convolutional coding structure contains two consecutive convolutional blocks, and each convolutional block includes convolution, batch normalization, and ReLU activation functions. For YCbCr images, the first convolutional block maps the input YCbCr image to an intermediate feature space and extracts the underlying color structure through nonlinear transformation; the second convolutional block enhances the output of the first convolutional block to a higher dimension, thereby improving semantic expressiveness. in, Image of YCbCr; As an intermediate feature; These are the high-dimensional color features of the YCbCr image.

3. The underwater salient target detection method based on multi-color space features according to claim 1, characterized in that: The high-dimensional color features of the YCbCr image and the high-dimensional color features of the HSV image are concatenated along the channel dimension and then injected into... The shallow layer of the encoder contains Y components of YCbCr images that carry rich local brightness structure information, and H and S components of HSV images that provide hue and saturation contrast cues. The fusion of the two can enhance the network’s ability to perceive blurred edges under degradation conditions and help the shallow layer capture the general outline of the target even when color information is insufficient. The high-dimensional color features of the YCbCr image are injected into... In the middle layer of the encoder, the Y component of the YCbCr image provides the framework of the regional structure, while the Cb and Cr components reflect the color consistency within the region. The high-dimensional color features of the HSV image are injected into Deeper into the encoder, the H component of the HSV image provides robust color category information for underwater color bias, enabling the distinction between the foreground and background when their overall colors are similar. The S component of the HSV image reflects the difference in color purity between the foreground and background. Under artificial lighting, the saturation of the foreground decreases due to increased white light, while the saturation of the background increases due to weakened lighting, providing additional criteria for deep semantic discrimination.

4. The underwater salient target detection method based on multi-color space features according to claim 1, characterized in that: for The encoder's first Original features of layer input If the first If the layer is shallow, then the first layer... Layer output features for: in, High-dimensional color features of the YCbCr image; High-dimensional color features of HSV images; Indicates concatenation of channel dimensions; This indicates that dynamic fusion is performed through a convolutional attention module and a gated fusion module; If the first If the layer is the middle layer, then the first layer... Layer output features for: If the first If the layer is deep, then the first... Layer output features for: 。 5. The underwater salient target detection method based on multi-color space features according to claim 1, characterized in that: The adaptive filtering of input color features by channel and spatial dimension through the convolutional attention module is specifically as follows: The convolutional attention module includes a channel attention module and a spatial attention module; The channel attention module is used to distinguish the effectiveness differences of different color component features. The channel attention module focuses on the input color features. Global average pooling and global max pooling are performed to aggregate global information and local most salient responses in the spatial dimension, respectively. The results of these two pooling methods are then fed into a shared multilayer perceptron (MLP) to generate channel weights. ; in, For activation functions; For global average pooling; This is for global max pooling; Channel weight With input color features Element-wise multiplication yields channel enhancement features. ; in, This represents element-wise multiplication; Channel enhancement features output by the channel attention module The spatial attention module performs average pooling and max pooling along the channel dimension, compressing multi-channel features into two single-channel feature maps, reflecting the average and maximum response intensities at each location, respectively. The two feature maps are then concatenated along the channel dimension and convolved with activation functions to generate spatial weights. ; in, For convolution; Indicates concatenation of channel dimensions; Spatial weight With channel enhancement features Element-wise multiplication yields the filtered color features. ; 。 6. The underwater salient target detection method based on multi-color space features according to claim 5, characterized in that: The process involves dynamically fusing the filtered color features with the original features of the current layer of the backbone network using a gated fusion module. The filtered color features are processed through convolution. Adapt to the original features of the current layer of the backbone network With the same channel dimension, the color features obtained after convolution are... ; The original features of the current layer of the backbone network Color features after convolution The weights are concatenated along the channel dimension and then generated through convolution and activation functions to produce gated weights. ; in, and These are the parameters and biases of the gated convolution, respectively; Gating weights For tensors with values ​​between 0 and 1, the dimension is the same as the original features of the current layer of the backbone network. same; In the form of residuals, the gating weights are... Weighted color features after convolution Then, the original features of the current layer of the backbone network are superimposed. The fused coding features are obtained. ; 。 7. The underwater salient target detection method based on multi-color space features according to claim 1, characterized in that: The The decoder uses and The encoder's symmetrical U-shaped structure gradually restores spatial resolution through step-by-step upsampling and feature stitching; Each decoding layer of the decoder first performs bilinear upsampling on the features of the previous layer to make its resolution consistent with the corresponding coding layer. Then, it concatenates the coded features fused with the previous layer along the channel dimension, and then performs feature fusion through convolutional blocks. After progressive upsampling and fusion through multiple decoding layers, the feature resolution is restored to the original input size. Finally, a 1×1 convolution maps the number of feature channels to 1, and the sigmoid activation function constrains the pixel values ​​to the [0,1] interval, outputting a saliency map of the same size as the input image. The closer the pixel value in the saliency map is to 1, the greater the probability that the position belongs to a salient target; the closer it is to 0, the greater the probability that it belongs to the background, thus obtaining the underwater salient target detection result.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

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

10. A computer program product comprising computer instructions, characterized in that: When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1 to 7.

Images