An edge-guided multi-attention RGBD underwater salient target detection method

Abstract

This invention discloses an edge-guided multi-attention RGBD underwater salient target detection method, comprising the following steps: S1, preprocessing the RGB image and depth image in the channel dimension and concatenating them to obtain an RMD image; S2, inputting the RMD image into a trained salient target detection model, the model including an encoder structure and a decoder structure, the encoder generating feature maps with progressively decreasing resolution based on the RMD image using a residual neural network, the decoder structure including an edge detection branch and a salient detection branch, the edge detection branch being used for progressively fusing feature maps, and the salient detection branch being used for progressively fusing feature maps and edge feature maps; S3, obtaining the salient image and edge image output by the salient target detection model as the target detection result. This invention preprocesses the input image according to the characteristics of underwater imaging, and simultaneously uses edge prediction maps to guide multi-level salient target detection, improving the edge details of the salient image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more particularly to an edge-guided multi-attention RGBD underwater salient target detection method. Background Technology

[0002] The salient object detection (SOD) task aims to detect the most noticeable objects in a scene and then extract pixel-level accurate target region masks. The advantage of salient object detection lies in its numerous applications, including foreground image cropping, foreground image evaluation, visual tracking, and image retrieval.

[0003] In recent years, the task of salient object detection has rapidly developed thanks to the ability of convolutional neural networks (CNNs) to extract deep semantic and shallow detail information at multiple scales. The development of new CNN-based architectures has also promoted the advancement of salient object detection, with the U-Net architecture being the most attractive due to its simple bottom-up structure capable of building rich feature maps. Attention mechanisms achieve a filtering effect by emphasizing important information and suppressing noise. Since the introduction of attention mechanisms, which detect the most noticeable objects in a scene, they have been widely used and have brought significant improvements to the task.

[0004] While significant progress has been made in salient object detection, research on underwater scenarios is scarce. The ocean is rich in resources, and marine ecology has high research value. Salient object detection can efficiently reduce redundant information in underwater images, improve the accuracy and recognition ability of underwater target detection, and effectively assist in the exploitation of marine resources and the conduct of marine ecological research. Therefore, underwater salient object detection is necessary. However, salient object detection faces challenges such as complex backgrounds and color deviations. To overcome these challenges, using depth maps to compensate for the spatial information missing in RGB images has become a new approach. The widespread use of depth sensors makes depth maps easier to obtain. Currently, methods using two-branch encoders to process RGB and depth maps separately have been proposed, which can effectively improve performance. However, if the performance improvement cannot compensate for the computational cost, introducing depth maps becomes redundant. Furthermore, to emphasize the semantic information of the channel dimension, channel attention mechanisms have been introduced, resulting in better model performance, but blurring still exists in edge regions. Summary of the Invention

[0005] In view of the shortcomings of existing technologies, this invention provides an edge-guided multi-attention RGBD underwater salient target detection method. This invention preprocesses the input image based on the characteristics of underwater imaging, and simultaneously utilizes an edge prediction map to guide deep-level salient target detection, thereby improving the detail of salient image edges.

[0006] The technical means employed in this invention are as follows:

[0007] An edge-guided multi-attention RGBD underwater salient target detection method includes the following steps:

[0008] S1. Obtain RGB image and depth image, perform channel-dimensional preprocessing on the RGB image and depth image and concatenate them to obtain RMD image;

[0009] S2. Input the RMD image into the trained salient object detection model, the salient object detection model including an encoder structure and a decoder structure.

[0010] The encoder generates feature maps with progressively decreasing resolution based on the RMD image using a residual neural network.

[0011] The decoder structure includes an edge detection branch and a salient detection branch.

[0012] The edge detection branch is used to fuse feature maps layer by layer to output an edge image.

[0013] The salient detection branch is used to fuse feature maps and edge feature maps layer by layer, and output a salient image with improved edge details;

[0014] S3. Obtain the salient image and edge image output by the salient target detection model as the target detection result.

[0015] Further, the RMD image is obtained by concatenating the RGB image and the depth image along the channel dimension, including:

[0016] Take the maximum value of the G and B channels in the RGB image as the M channel;

[0017] The R channel, M channel, and depth map are concatenated along the channel dimension to form a new input RMD map.

[0018] Furthermore, the encoder structure includes:

[0019] The preprocessing layer consists of a 7×7 convolutional layer and a 2×2 max pooling layer to obtain feature maps with a resolution reduced to 1 / 4 of the original size.

[0020] The basic convolutional block consists of three convolutional layers, each using batch normalization and the ReLU activation function;

[0021] Cross-layer connections are added between the input and output of each basic convolutional block to avoid gradient vanishing or gradient exploding.

[0022] Combining basic convolutional blocks and cross-layer connections into a single residual neural module, the output feature map has multiple scales from shallow to deep levels.

[0023] Furthermore, the edge detection branch includes an edge pyramid convolution module and an edge detection branch spatial attention module;

[0024] The edge pyramid convolution module inputs the deepest level feature map into the multi-scale convolution module, and then inputs the outputs of multiple convolutions at different scales into the convolution module to output the edge feature map.

[0025] The edge detection branch spatial attention module receives an edge feature map and a feature map, upsamples the small-scale edge feature map to obtain the same scale as the large-scale feature map, and concatenates the edge feature map and the feature map along with their pixel-level dot product matrix in the channel dimension to obtain the concatenated feature map.

[0026] The concatenated feature maps are converted into spatial attention weights, which are used to calibrate the spatial features of the concatenated feature maps, resulting in edge attention maps.

[0027] First, the obtained edge attention map is residually connected with the original feature map and the upsampled edge feature map respectively. Then, the two branch channel dimensions are concatenated and convolutional operation is performed. Finally, the edge feature map is output.

[0028] Furthermore, the salient detection branch includes: a scale pyramid convolution module and a salient detection branch spatial attention module;

[0029] The scale pyramid convolution module takes the deepest level feature map as input and convolves it at different scales. Then, it inputs the outputs of multiple convolutions at different scales into the convolution and outputs the edge feature map.

[0030] The salient detection branch spatial attention module receives an edge feature map, a salient feature map, and a feature map. It generates an edge attention map for calibrating the spatial features of the salient feature map by normalizing the edge feature map using a function. The calibrated salient feature map is then concatenated with the edge feature map along the channel dimension to obtain an edge-enhanced salient feature map. This map generates three tensors—query, key, and value—through convolution and deformation, which are then input into the self-attention module. The self-attention module outputs an attention feature map.

[0031] The attention feature map and the feature map, along with their dot product matrix, are concatenated along the channel dimension to obtain the concatenated feature map. This concatenated feature map is then normalized using a function to generate a semantic attention map for calibrating the channel semantic features. Finally, the semantically enhanced feature map is obtained by multiplying the feature map pixels by the semantic attention map.

[0032] The semantically enhanced feature map is residually connected to the original feature map and the attention feature map, and then convolutional is performed to obtain the final salient feature map.

[0033] Furthermore, the training steps of the salient target detection model include:

[0034] Acquire training set data, which includes land images and underwater images, as well as images of salient targets in Ground Truth;

[0035] The network was trained using the aforementioned land and underwater images as input data, respectively.

[0036] Calculate the loss value between the salient target image output by the network and the salient target image in the GroundTruth training set, perform error backpropagation based on the loss value, and update the weights of the network.

[0037] Further, the loss value between the salient target image output by the network and the salient target images in the Ground Truth training set is calculated, and error backpropagation is performed based on the loss value to update the weights of the network, including:

[0038] For each level of the edge detection branch and the saliency detection branch, the loss value is calculated relative to the real salient image and the edge image, respectively. The loss value is minimized using stochastic gradient descent to obtain the optimal model.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. This invention uses a concatenation method and introduces a preprocessing module when testing underwater datasets to make the input more consistent with the underwater scene and adapt to the characteristics of underwater lighting.

[0041] 2. This invention proposes a multi-level edge progressively guided saliency detection method, which solves the problem of information loss that may occur in methods guided by different resolutions.

[0042] 3. Considering the more significant spatial features of edge maps, this invention designs a spatial attention mechanism. This mechanism suppresses noisy edges and emphasizes real edges by weighting the pixels of the edge map, making the rough edge map more refined, which is beneficial for accurately guiding the detection of significant targets. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram of the salient target detection model structure in an embodiment of the present invention.

[0045] Figure 2 This is a structural diagram of the edge pyramid convolution module in an embodiment of the present invention.

[0046] Figure 3 This is a structural diagram of the edge prediction branch spatial attention module in an embodiment of the present invention.

[0047] Figure 4 This is a structural diagram of the multi-scale pyramid convolution module in an embodiment of the present invention.

[0048] Figure 5 This is a structural diagram of the attention module for significant detection of branch spatial channels in an embodiment of the present invention.

[0049] Figure 6 This is a schematic diagram of the training process of the salient target detection model in an embodiment of the present invention. Detailed Implementation

[0050] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] This invention provides an edge-guided multi-attention RGBD underwater salient target detection method, comprising the following steps:

[0052] S1. Obtain RGB image and depth image, and concatenate the RGB image and depth image along the channel dimension to obtain RMD image.

[0053] Specifically, the underwater image preprocessing module, considering the characteristics of underwater light, takes the maximum value of the G and B channels in the RGB image as the M channel, and concatenates the R channel, M channel, and depth map along the channel dimension to form a new input RMD image, which can be described as follows:

[0054] M = Max(G, R)(1)

[0055] RMD = £(R,M,Depth)(2)

[0056] Where Max represents taking the maximum value, and £ represents concatenating the channel dimensions.

[0057] S2. Input the RMD image into the trained salient object detection model. The salient object detection model includes an encoder structure and a decoder structure. The encoder generates feature maps with progressively decreasing resolution based on the RMD image using a residual neural network. The decoder structure includes an edge detection branch and a salient detection branch. The edge detection branch is used to fuse the feature maps layer by layer and output an edge image. The salient detection branch is used to fuse the feature maps and edge feature maps layer by layer and output a salient image with enhanced edge details. Specifically:

[0058] Furthermore, the encoder uses a residual neural network as the backbone network to generate feature maps with progressively decreasing resolution, mainly including: preprocessing layers, basic convolutional blocks, and cross-layer connections. Finally, several basic convolutional blocks and cross-layer connections are combined into a residual neural module.

[0059] Preferably, the preprocessing layer consists of a 7×7 convolution, a batch normalization layer, a ReLU activation function, and a 2×2 max pooling layer to obtain a feature map Res1 with a resolution reduced to 1 / 4 of the original size.

[0060] Preferably, the basic convolutional block consists of three convolutional layers: the first convolutional layer consists of a 1×1 convolution, a batch normalization layer, and a ReLU activation function; the second convolutional layer consists of a 3×3 convolution with a stride of 2, a batch normalization layer, and a ReLU activation function, wherein setting the stride to 2 has the function of downsampling; and the third convolutional layer consists of a 1×1 convolution and a batch normalization layer.

[0061] Preferably, cross-layer connections are added between the input and output of each basic convolutional block to avoid gradient vanishing or gradient explosion. The input of the basic convolutional block passes through a downsampling layer and is cross-connected with the output of the basic convolutional block. After the connection, the output of the cross-layer connection is obtained by passing through the ReLU activation function. The downsampling layer consists of a 1×1 convolution with a stride of 2 and a batch normalization layer.

[0062] Preferably, several basic convolutional blocks and cross-layer connections are combined into a single residual neural module. Taking ResNet-50 as an example, the first residual neural module consists of 3 basic convolutional blocks and cross-layer connections, and outputs feature map Res2; the second residual neural module consists of 4 basic convolutional blocks and cross-layer connections, and outputs feature map Res3; the third residual neural module consists of 6 basic convolutional blocks and cross-layer connections, and outputs feature map Res4; and the fourth residual neural module consists of 3 basic convolutional blocks and cross-layer connections, and outputs feature map Res5. Thus, the output has feature maps Res with hierarchical levels from shallow (large scale) to deep (small scale). i ,i={1,2,3,4,5}.

[0063] Furthermore, the decoding stage uses a dual-branch approach of edge detection and saliency detection, with the edge detection branch being gradually integrated into the saliency detection branch.

[0064] Preferably, starting from the fourth level, the edge feature map output from the deep level is upsampled and input into the spatial attention module of the edge prediction branch of the shallow level along with the residual neural network feature map of the current level, and the edge feature map of the current level is output. The edge feature map output from the deep level and the salient feature map output from the deep level are input into the spatial attention module of the salient detection branch, and the spatial attention feature map is output. The spatial attention feature map and the residual neural network feature map of the current level are input into the dot product attention module of the salient detection branch, and the salient feature map of the current level is output.

[0065] Preferably, in order to learn deep-level multi-scale edge feature representations, this invention designs an edge pyramid convolution module, such as... Figure 2 As shown, this module uses convolutions of different sizes to extract edge features. First, the deepest layer feature map, Res5, is input, then it undergoes convolutions of different sizes to obtain different outputs y. i If i = {1, 2, 3, 4, 5}, then concatenate the channels, and finally pass through a 1×1 convolution, a batch normalization layer, and a ReLU activation function to output the edge feature map edge4, which can be described as follows:

[0066]

[0067] edge4 = ReLU(Bn(conv) 1×1 (￡(y1,y2,y3,y4,y5)))) (4)

[0068] Where mean1 represents the average value along the channel dimension, and conv 1×1 conv represents a 1×1 convolution. 3×3 conv represents a 3×3 convolution. 1×3 conv represents a 1×3 convolution. 3×1 represents a 3×1 convolution, Bn represents a batch normalization layer, and ReLU represents the ReLU activation function.

[0069] Preferably, in order to extract more accurate edge feature maps, this invention designs a novel edge prediction branch spatial attention module, such as... Figure 3 As shown.

[0070] a. First, generate a small-scale edge feature map. i+1 Upsampled to large-scale feature maps Res i At the same scale, edge feature maps are then concatenated along the channel dimension. i+1 and feature map Res iAnd the pixel-level dot product matrix of the two is used to obtain the concatenated feature map edge. i+ 1 can be expressed as:

[0071] edge sa =(up(edge) i+1 ),Res i ,up(edge i+1 Res i (5)

[0072] Where o represents pixel-level matrix dot product, and ￡ represents concatenation of channel dimensions.

[0073] b. Secondly, the concatenated feature map edge sa The average value is taken along the channel dimension, followed by convolution and sigmoid function normalization to convert it into spatial attention weights, which are used to calibrate the edges of the concatenated feature maps. sa Spatial features, spatial attention weight matrix, dot product feature map edge sa Obtain the edge attention map. spa It can be expressed as:

[0074] edge spa =edge sa osigmoid(conv(mean1(edge sa (6)

[0075] Where o represents pixel-level matrix dot product, conv represents 3×3 convolution, and mean1 represents the average value across the channel dimension.

[0076] c. Finally, first process the obtained edge attention map... spa Compare with the original feature map Res respectively i Upsampled edge feature map i +1 is used for residual connection, then the two branches are concatenated by channel dimensions and then convolutional to obtain the final edge feature map. i It can be expressed as:

[0077] edge i =conv(￡(conv(Res) i +conv(edge spa )),

[0078] conv(up(edge i+1 )+conv(edge spa )))) (7)

[0079] Where ￡ represents concatenation of channel dimensions, and conv represents 3×3 convolution.

[0080] Preferably, in order to learn deep-level multi-scale semantic feature representations, this invention designs a multi-scale pyramid convolution module, such as... Figure 4 As shown, this module uses convolutions of different sizes to extract semantic features. First, the deepest level feature map, Res5, is input, then it undergoes convolutions of different sizes to obtain different outputs y. i Let i = {1, 2, 3, 4, 5}, then concatenate the channels, and finally pass through conv. 1×1 The convolutional output feature map f4 can be expressed as:

[0081]

[0082] f4 = ReLU(Bn(conv) 1×1 (￡(y1,y2,y3,y4,y5)))) (9)

[0083] Where mean 2,3 conv represents taking the average of the second and third dimensions. 1×1 conv represents a 1×1 convolution. 5×5 conv represents a 5×5 convolution. 9×9 conv represents a 9×9 convolution. 13×13 This represents a 13×13 convolution, Bn represents a batch normalization layer, and ReLU represents the ReLU activation function.

[0084] Preferably, to effectively fuse salient feature maps and edge feature maps, this invention designs a new salient feature and edge feature fusion module; to extract more accurate salient region location information, this invention uses a spatial channel attention module. The above modules are as follows: Figure 5 As shown.

[0085] a. First, normalize the edge feature map using the sigmoid function. i+1 Generate an edge attention map for calibrating the spatial features of the salient feature map, matrix dot product of the salient feature map f. i+1 Residual connections are then performed to obtain the calibrated salient feature map. The calibrated salient feature map is then compared with the edge feature map. i+1 By concatenating the data along the channel dimension, we obtain the salient feature map fe for edge enhancement, which can be expressed as:

[0086] fe = (sigmoid(edge) i+1 )f i+1 +f i+1 ,edge i+1 (10)

[0087] Where o represents pixel-level matrix dot product, and ￡ represents concatenation of channel dimensions.

[0088] b. Then, the query fe is generated from the salient feature map fe through convolution and deformation. q , keyfe k Sum of values ​​fe v The three tensors can be expressed as:

[0089]

[0090] Here, `reshape` represents a deformation operation that transforms the height and width dimensions to a single dimension, i.e., converting a feature map of shape [B,C,H,W] into a feature map of shape [B,C,H×W]. `conv`... q ,conv k ,conv v This represents three 3×3 convolutions.

[0091] Input the above three tensors into the dot product attention module, and input the query fe q , keyfe k Sum of values ​​fe v Three tensors, query fe q and key fe k The attention matrix, with value fe, is obtained by multiplying two tensors using matrix multiplication and the softmax function. v The attention matrix is ​​multiplied by matrix multiplication and then residually concatenated with the salient feature map fe to obtain the attention feature map f. att The dot product attention module can be described as follows:

[0092] attention = softmax(fe) q ×(fe k ) T (12)

[0093] f att =f v ×attention+fe (13)

[0094] in(*) T This indicates the transpose operation.

[0095] c. Secondly, first concatenate the attention feature maps f along the channel dimension. att and feature map Res i The dot product of these two matrices yields the concatenated feature map f. ca Then, the concatenated feature map f is normalized by the sigmoid function. ca Generate a semantic attention map for calibrating channel semantic features, then feature map f ca Pixel-to-semantic attention map multiplication yields semantically enhanced feature map f gap It can be expressed as:

[0096] f ca =(up(f att ),Res i ,up(f att Res i (14)

[0097] f gap =f ca osigmoid(conv(mean 2,3 (fa))) (15)

[0098] Where o represents pixel-level matrix dot product, ￡ represents concatenation of channel dimensions, and mean 2,3 This indicates that the average value is taken in the height and width dimensions.

[0099] d. Finally, first process the semantically enhanced feature map f gap Compare with the original feature map Res respectively i Attention feature map f att Perform residual connections, then concatenate the two branch channel dimensions and perform convolution to obtain the final feature map f. i It can be expressed as:

[0100]

[0101] Where ￡ represents concatenation of channel dimensions, and conv represents 3×3 convolution.

[0102] A salient target detection model was constructed using the above methods. To obtain the optimal detection model, it needs to be trained. This invention uses both land and underwater datasets to train the proposed underwater salient target detection network model. Since there is currently limited research on underwater salient target detection, a land dataset was chosen for comparison with other land-based methods to verify the model's performance. To improve the model's generalization ability, such as learning from clear images of people on land to improve the accuracy of detecting divers, a combination of a land training set and an underwater test set was chosen. Figure 6 As shown, the specific training steps are as follows:

[0103] The labels operator was used to extract target edge images from the real salient images in the dataset as ground truth edge images. Two training strategies were employed during the training phase: training using the land dataset and training using a hybrid dataset, which included both land and underwater datasets. For the land dataset, 700, 1500, and 785 samples were selected from the NLPR, NJUD, and DUT-D datasets, respectively, totaling 2985 samples for training; the remaining samples were used for testing. For the underwater dataset, 2051 samples were selected from the USOD10k dataset as the training set, with the remaining samples used for testing.

[0104] During training, a warm-up and linear decay strategy is used to adjust the learning rate. The maximum learning rate for the backbone is 0.05. The momentum and decay rate are set to 0.9 and 5e-4, respectively. The loss between the output of each level of the two branches and the real image is calculated. The loss value is minimized using stochastic gradient descent to obtain the optimal model. Finally, the predicted salient image and edge image are output during the testing phase.

[0105] As a supervised training method, the classic binary cross-entropy (BCE) function is used as the loss function in the salient object detection branch. Let y s , Given the true saliency map and the predicted saliency map, respectively, the binary cross-entropy function (BCE) can be expressed as:

[0106]

[0107] In addition, to obtain better edge performance, the Intersection over Union (IoU) ratio is used as an additional loss function. The IoU ratio can be expressed as:

[0108]

[0109] In the edge detection branch, the binary cross-entropy function (BCE) is used as the loss function, denoted as y. e , Given the true edge map and the predicted edge map, respectively, the binary cross-entropy function (BCE) can be expressed as:

[0110]

[0111] S3. Obtain the salient image and edge image output by the trained salient object detection model as the object detection result.

[0112] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An edge-guided multi-attention RGBD underwater salient object detection method, characterized in that, The method comprises the following steps: S1, acquiring an RGB image and a depth image, performing channel dimension preprocessing on the RGB image and the depth image, and acquiring an RMD image in series, including taking the maximum value of the G channel and the B channel in the RGB image as the M channel, and concatenating the R channel, the M channel and the depth image in the channel dimension as a new input RMD image; S2, inputting the RMD image into a trained salient target detection model, the salient target detection model comprising an encoder structure and a decoder structure, the encoder generates feature maps with decreasing resolution level by level based on a residual neural network according to the RMD image, the decoder structure comprises an edge detection branch and a salient detection branch, the edge detection branch is used for fusing feature maps level by level, and outputs an edge image, the salient detection branch is used for fusing feature maps and edge feature maps layer by layer, and outputs a salient image with enhanced edge details; S3, acquiring the salient image and the edge image output by the salient target detection model as a target detection result.

2. The edge-guided multi-attention RGBD underwater salient object detection method according to claim 1, wherein, The encoder structure comprises: a pre-processing layer including one convolutional layer and one max-pooling layer to obtain a feature map with a resolution reduced to 1 / 4 of the original size; a basic convolution block comprising three convolution layers, each convolution layer using batch normalization and a ReLU activation function; a cross-layer connection added between the input and output of each basic convolution block to avoid gradient disappearance or gradient explosion; combining the basic convolution block and the cross-layer connection as a residual neural module, and the output feature map has multiple scales from a shallow level to a deep level.

3. The edge-guided multi-attention RGBD underwater salient object detection method according to claim 1, wherein, The edge detection branch comprises an edge pyramid convolution module and an edge detection branch spatial attention module; The edge pyramid convolution module inputs the deepest level feature map into a multi-scale convolution module, and then inputs the outputs of multiple different scale convolutions into convolution, output edge feature map; The edge detection branch spatial attention module receives edge feature maps and feature maps, up-samples small-scale edge feature maps to obtain the same scale as large-scale feature maps, concatenates the edge feature maps and the feature maps in the channel dimension, and pixel-level point multiplies the matrices to obtain concatenated feature maps, converts the concatenated feature maps into spatial attention weights for calibrating the spatial features of the concatenated feature maps to obtain an edge attention map, firstly, the obtained edge attention map is respectively connected with the original feature map and the up-sampled edge feature map in residual, then the two branch channel dimensions are concatenated and convolution operation is performed, and finally the edge feature map is output.

4. The edge-guided multi-attention RGBD underwater salient object detection method according to claim 1, wherein, The salient detection branch comprises a scale pyramid convolution module and a salient detection branch spatial attention module; The scale pyramid convolution module inputs a deepest level feature map to different scale convolutions, and then inputs outputs of the different scale convolutions to a pooling layer convolution, output edge feature map The salient detection branch spatial attention module receives edge feature maps, salient feature maps and feature maps, normalizes the edge feature maps by a function to generate an edge attention map for calibrating the spatial features of the salient feature maps, concatenates the calibrated salient feature maps and the edge feature maps in the channel dimension to obtain edge-enhanced salient feature maps, generates three tensors of query, key and value through convolution and deformation, and inputs them into a self-attention module, and outputs an attention feature map through the self-attention module, concatenates the attention feature map and the feature map in the channel dimension and pixel multiplies the matrices to obtain concatenated feature maps, and normalizes the concatenated feature maps by a function to generate a semantic attention map for calibrating the channel semantic features, and pixel multiplies the semantic attention map with the semantic-enhanced feature map to obtain a semantic-enhanced feature map. The semantic enhanced feature map is respectively connected with the original feature map and the attention feature map in a residual manner, and a convolution operation is performed to obtain a final salient feature map.

5. The edge-guided multi-attention RGBD underwater salient object detection method according to claim 1, wherein, The training step of the salient object detection model comprises: Obtaining training set data, wherein the training set data comprises land images and underwater images, and salient object images of GroundTruth; Training the network by taking the land images and the underwater images as input data of the network respectively; Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network.

6. The edge-guided multi-attention RGBD underwater salient object detection method according to claim 5, characterized in that, Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network and the salient object images of GroundTruth in the training set, performing error back propagation according to the loss value, and updating the weight of the network, comprises: Calculating a loss value between the salient object images output by the network

Images