A method for detecting a salient object based on a multi-scale dilated convolutional neural network

Abstract

This invention discloses a salient object detection method based on a multi-scale dilated convolutional neural network. The method includes: extracting multi-scale features from the input image; inputting the multi-scale features into a dilated residual convolutional module to obtain fused features including contextual information of the multi-scale features; inputting the fused features into multiple channel attention modules to obtain multiple salient features; performing dimensionality reduction activation on each salient feature to generate a saliency map; and performing deep supervised training using a hybrid loss function that combines cross-entropy and cross-union loss. The method of this invention, based on a multi-scale dilated convolutional neural network, fully captures rich global and local semantic information in the image by using a dilated residual convolutional module, solving the problem of shallow encoder depth and insufficient information extraction. Simultaneously, the designed channel attention modules enable the network to focus on the target region, effectively improving the accuracy of object detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more specifically to a salient object detection method based on a multi-scale dilated convolutional neural network. Background Technology

[0002] Visual saliency refers to the most prominent and unique salient regions in an image that differ from the background. Salient object detection is a crucial preprocessing step in computer vision tasks, aiming to extract useful features from a given image, detect visually salient objects or regions, and output a corresponding saliency map. In recent years, salient object detection has been widely researched and applied, such as in image translation, semantic segmentation, object recognition, and object tracking.

[0003] Saliency detection methods can be broadly categorized into two types: traditional saliency detection methods and deep learning-based saliency detection methods. Traditional saliency detection methods primarily rely on human intuition or heuristic priors, such as color comparison, background comparison, boundary point priors, and texture, to manually extract features for target detection. However, due to the influence of complex environments, it is difficult to accurately locate and segment target regions using high-level semantic information.

[0004] In recent years, the rapid development of deep learning technology has led to the widespread proposal of deep learning-based salient object detection methods, which have significantly improved performance compared to traditional methods. Deep convolutional neural networks can automatically extract multi-level and multi-scale features, including low-level features rich in edge information and high-level features containing global semantic information, thereby accurately locating and completely segmenting salient objects. However, existing deep learning-based salient object detection methods still have shortcomings. In challenging scenarios, such as complex backgrounds, highly similar salient objects to the background, and low-contrast scenes, model performance needs improvement. Some methods, such as the CPD method proposed by WU et al., directly concatenate convolution and dilated convolution operations with different receptive fields. While this optimizes multi-scale features and increases feature diversity, it fails to fully utilize some feature information, and the captured information is not correlated, reducing the model's detection performance. Other methods employ attention mechanisms or selectively highlight salient object regions through gating functions to reduce interference from redundant information. However, attention mechanisms based on the same scale have limitations for salient objects with complex types and unpredictable sizes. Summary of the Invention

[0005] In view of this, the present invention provides a salient target detection method based on a multi-scale dilated convolutional neural network, which can solve the technical problem of inaccurate detection results caused by insufficient extraction of high-level and low-level feature information and insufficient correlation mining between them in complex scenes.

[0006] To solve the above-mentioned technical problems, the present invention is implemented as follows.

[0007] A salient object detection method based on a multi-scale dilated convolutional neural network includes:

[0008] Step S1: Use the RGB image to be processed as the input image, input the input image into the pre-trained ResNet34 network model, and extract the multi-scale features of the input image;

[0009] Step S2: Input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input into the second convolutional layer, and each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module;

[0010] Step S3: Input the fused features into multiple channel attention modules respectively to obtain multiple saliency features;

[0011] Step S4: Dimensionally reduce and activate each salient feature to generate a saliency map, and perform deep supervised training using a hybrid loss function that combines cross-entropy and cross-union loss.

[0012] Preferably, the channel attention module includes an average pooling module, a max pooling module, a first convolutional branch, a second convolutional branch, and a third convolutional module. The fusion feature is input into the average pooling module and the max pooling module respectively. The results processed by the average pooling module and the max pooling module are fused to obtain a first fusion feature. The first fusion feature is input into the first convolutional branch and the second convolutional branch respectively. The first fusion feature, the processing result obtained by the first convolutional branch, and the processing result obtained by the second convolutional branch are fused to obtain a second fusion feature. The second fusion feature is input into the third convolutional module to obtain a saliency feature.

[0013] Preferably, in step S2, wherein:

[0014] After performing dilated convolution, both the first and second dilated convolution modules undergo batch normalization and ReLU nonlinear activation operations; the fill rate of the first dilated convolution module is d0 = {2, 4, 6}, and the fill rate of the second dilated convolution module is d1 = {1, 2, 3}.

[0015] The output of the branch is F. i ,

[0016] F i =∑(D i0 (res(f j )), D i1 (res(f j )), f j )

[0017] Where i represents the number of branches containing residual modules and dilated convolutions, j represents the number of multi-scale feature branches, and f j The multi-scale dimensionality reduction features are represented by res(·), which represents the convolution operation performed by the residual module. i0 D i1 ...

[0018] The output F of each dilated convolution branch i Inputting the second convolutional layer yields a fused feature Y that includes contextual information from the multi-scale features. j =Conv(C(F0, F1, F2, f j )), where Conv represents dimensionality reduction convolution operation and C represents channel-dimensional feature fusion operation.

[0019] Preferably, in step S3, wherein:

[0020] The first convolutional branch and the second convolutional branch have different kernel sizes. The first convolutional branch includes a 3×3 convolutional layer, a fully connected layer, and a sigmoid layer. The first fusion feature is input into the first convolutional branch to obtain the first influence α of different channel levels on saliency information. The second convolutional branch includes a 5×5 convolutional layer, a fully connected layer, and a sigmoid layer. The first fusion feature is input into the second convolutional branch to obtain the second influence β of different channel levels on saliency information. The first influence α, the second influence β, and the first fusion feature ε are fused in the following manner:

[0021]

[0022] in, M represents the product of coefficients. jThis represents the saliency feature of the channel attention module output of the multi-scale branch j.

[0023] Preferably, in step S4, the hybrid loss function L is

[0024] L=λ1L BCE (P, G) + λ²L IOU (P, G)

[0025] Where λ1 and λ2 are the weight coefficients of the cross-entropy loss function BCE and the IOU loss function, respectively, and L BCE (P, G) represents the cross-entropy loss function of the saliency plot, L IOU (P, G) represents the IOU loss function for the saliency map, where P is the predicted result of the saliency map and G is the ground truth label of the saliency map.

[0026]

[0027]

[0028] Where l represents the l-th pixel in the input image, n represents the total number of pixels in the input image, and G l P represents the truth label of the saliency map for the l-th pixel. l This represents the prediction result of the saliency map for the l-th pixel.

[0029] The present invention provides a salient object detection device based on a multi-scale dilated convolutional neural network, the device comprising:

[0030] Multi-scale encoding module: configured to take the RGB image to be processed as the input image, input the input image into a pre-trained ResNet34 network model, and extract the multi-scale features of the input image;

[0031] Dilated Residual Convolution Module: Configured to input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input to the second convolutional layer, each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module;

[0032] The salient feature acquisition module is configured to input the fused features into multiple channel attention modules respectively to obtain multiple salient features;

[0033] The saliency map generation module is configured to perform dimensionality reduction activation on each saliency feature to generate a saliency map, and uses a hybrid loss function that combines cross-entropy and cross-union loss for deep supervised training.

[0034] The present invention provides a computer-readable storage medium storing a plurality of instructions; the plurality of instructions are used by a processor to load and execute the method as described above.

[0035] The present invention provides an electronic device, characterized in that the electronic device comprises:

[0036] A processor is used to execute multiple instructions;

[0037] Memory, used to store multiple instructions;

[0038] The plurality of instructions are to be stored in the memory and loaded and executed by the processor as described above.

[0039] Beneficial effects:

[0040] (1) The method of the present invention is based on a multi-scale dilated convolutional neural network and uses a pre-trained ResNet34 as a feature encoder to effectively extract multi-scale initial features. By using the dilated residual convolution module, it fully captures the rich global and local semantic information in the image, thus solving the problem of shallow encoder depth and insufficient information extraction.

[0041] (2) By designing a channel attention module, the network can be further focused on the salient target area, reducing the interference of redundant information.

[0042] (3) A hybrid loss function is adopted, which combines cross-entropy loss and cross-proportion loss to effectively improve the accuracy of image salient target detection.

[0043] (4) The end-to-end deep supervision training method has the advantages of fast training speed, small amount of computation, strong versatility and high detection accuracy.

[0044] (5) This invention effectively utilizes multi-scale feature information to fully explore the correlation between high-level and low-level information. Attached Figure Description

[0045] Figure 1 A schematic diagram of the salient target detection method based on multi-scale dilated convolutional neural networks provided by the present invention;

[0046] Figure 2 A schematic diagram of the salient object detection architecture based on a multi-scale dilated convolutional neural network provided by the present invention;

[0047] Figure 3 This is a schematic diagram of the structure of the dilated residual convolution module provided by the present invention;

[0048] Figure 4 A schematic diagram of the channel attention module structure provided by the present invention;

[0049] Figure 5 This is a schematic diagram of the device structure for the salient target detection method based on a multi-scale dilated convolutional neural network provided by the present invention. Detailed Implementation

[0050] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0051] like Figures 1-2 As shown, this invention proposes a salient target detection method based on a multi-scale dilated convolutional neural network, comprising the following steps:

[0052] Step S1: Use the RGB image to be processed as the input image, input the input image into the pre-trained ResNet34 network model, and extract the multi-scale features of the input image;

[0053] Step S2: Input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input into the second convolutional layer, and each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module;

[0054] Step S3: Input the fused features into multiple channel attention modules respectively to obtain multiple saliency features;

[0055] Step S4: Dimensionally reduce and activate each salient feature to generate a saliency map, and perform deep supervised training using a hybrid loss function that combines cross-entropy and cross-union loss.

[0056] Furthermore, the channel attention module includes an average pooling module, a max pooling module, a first convolutional branch, a second convolutional branch, and a third convolutional module. The fusion feature is input into the average pooling module and the max pooling module respectively. The results processed by the average pooling module and the max pooling module are fused to obtain a first fusion feature. The first fusion feature is input into the first convolutional branch and the second convolutional branch respectively. The first fusion feature, the processing result obtained by the first convolutional branch, and the processing result obtained by the second convolutional branch are fused to obtain a second fusion feature. The second fusion feature is input into the third convolutional module to obtain a saliency feature.

[0057] In step S1, the following is stated:

[0058] In this embodiment, a ResNet34 network model pre-trained on ImageNet is used as the encoder to extract multi-scale features I from the input image. j Taking j = 0, 1, 2, 3, 4 as an example, conv1, conv2_3, conv3_4, conv4_6, and conv5_3 are selected as the five output layers of the encoder, respectively; the input image I(H,W,3) is downsampled to generate feature I. j (j = 0, 1, 2, 3, 4), where H, W, 3 are the height, width, and number of channels of the feature, respectively. Then the output feature of conv1 is I0(H, W, 64), the output feature of conv2_3 is I1(H / 2, W / 2, 64), the output feature of conv3_4 is I2(H / 4, W / 4, 128), the output feature of conv4_6 is I3(H / 8, W / 8, 256), and the output feature of conv5_3 is I4(H / 16, W / 16, 512).

[0059] Step S2, as follows Figure 3 As shown, where:

[0060] The first convolutional layer is used to reduce the dimensionality of the multi-scale features.

[0061] After performing dilated convolution, both the first and second dilated convolution modules undergo batch normalization and ReLU nonlinear activation operations; the fill rate of the first dilated convolution module is d0 = {2, 4, 6}, and the fill rate of the second dilated convolution module is d1 = {1, 2, 3}.

[0062] The output of the branch is F. i ,

[0063] F i =∑(D i0 (res(f j)), D i1 (res(f j )), f j ), (i=0, 1, 2; j=0, 1, 2, 3, 4)

[0064] Where i represents the number of branches containing residual modules and dilated convolutions, j represents the number of multi-scale feature branches, and f j The multi-scale dimensionality reduction features are represented by res(·), which represents the convolution operation performed by the residual module. i0 D i1 ...

[0065] The output F of each dilated convolution branch i Inputting the second convolutional layer yields a fused feature Y that includes contextual information from the multi-scale features. j =Conv(C(F0, F1, F2, f j )), where Conv represents dimensionality reduction convolution operation and C represents channel-dimensional feature fusion operation.

[0066] This invention utilizes multi-kernel dilated convolution and residual convolution to fully extract multi-scale contextual information, enhancing the diversity and relevance of features. Through residual convolution and dilated convolution with different receptive fields, as well as information fusion, it avoids information loss, captures more salient information in the context, and makes the features more generalizable.

[0067] For example, using the multi-scale features as input to the dilated residual convolution module, the input features f are first reduced in dimensionality through a 1×1 convolution operation. j Secondly, the data is fed into three sets of convolutional operations consisting of a classic residual block and two dilated convolutions, respectively. Each dilated convolution is followed by batch normalization and ReLU nonlinear activation, with padding rates of d0 = {2, 4, 6} and d1 = {1, 2, 3}. Thirdly, all features output from the dilated convolutions are compared with f... j Feature fusion is performed to obtain the feature F of each branch. i , of which F i The calculation formula is:

[0068] F i =∑(D i0 (res(f j )), D i1 (res(f j )), f j ), (i=0, 1, 2; j=0, 1, 2, 3, 4)

[0069] Where res(·) represents the residual block, D i0D i1 ...

[0070] Finally, dimensionality reduction is performed using 1×1 convolution operations to obtain feature Y with deep contextual semantic information. j The calculation formula is as follows:

[0071] Υ j =Conv(C(F0, F1, F2, f j ))

[0072] Step S3, as follows Figure 4 As shown, where:

[0073] For the channel attention module, the kernel sizes of the first and second convolutional branches are different. The first convolutional branch includes a 3×3 convolutional layer, a fully connected layer, and a sigmoid layer. The first fusion feature is input into the first convolutional branch to obtain the first influence α of different channel levels on saliency information. The second convolutional branch includes a 5×5 convolutional layer, a fully connected layer, and a sigmoid layer. The first fusion feature is input into the second convolutional branch to obtain the second influence β of different channel levels on saliency information. The first influence α, the second influence β, and the first fusion feature ε are fused in the following manner:

[0074]

[0075] in, M represents the product of coefficients. j This represents the saliency feature output by the channel attention module of the multi-scale branch j. The saliency feature is the result of fusing the features of the two branches with ε.

[0076] This invention enhances the saliency of features across multiple channels using a channel attention module, avoiding interference from redundant information. By strengthening the saliency of information within channels, interference from redundant background information is mitigated to some extent, making the features more robust.

[0077] For example, feature Y j As input to the channel attention module, average pooling and max pooling are used to retain different information in the features. After fusion, the resulting feature ε is processed by two branches of convolution operations with 3×3 and 5×5 kernels, cascaded fully connected operations, and sigmoid activation function operations to obtain the degree of influence α and β of different channel levels on saliency information. These are then multiplied with ε and fused with features. Finally, through a 1×1 convolution operation, a more clustered saliency information feature is obtained. The formula can be expressed as:

[0078]

[0079] in, M represents the product of coefficients. j This represents the result after fusing the features of the two branches with ε.

[0080] In step S4, for each significant feature M... j Dimensionality reduction activation is performed to generate a saliency map, and deep supervised training is carried out using a hybrid loss function that combines cross-entropy and cross-union loss.

[0081] The hybrid loss function L is

[0082] L=λ1L BCE (P, G) + λ²L IOU (P, G)

[0083] Where λ1 and λ2 are the weight coefficients of the cross-entropy loss function BCE and the IOU loss function, respectively, and L BCE (P, G) represents the cross-entropy loss function of the saliency plot, L IOU (P, G) represents the IOU loss function for the saliency map, where P is the predicted result of the saliency map and G is the ground truth label of the saliency map.

[0084]

[0085]

[0086] l represents the l-th pixel in the input image, n represents the total number of pixels in the input image, and G l P represents the truth label of the saliency map for the l-th pixel. l This represents the prediction result of the saliency map for the l-th pixel.

[0087] Furthermore, using the hybrid loss function as the optimization objective, an end-to-end training method is adopted to train the multi-scale dilated convolutional neural network composed of the dilated residual convolutional module and the multiple channel attention modules, thereby effectively locating and segmenting the target and obtaining an accurate saliency map.

[0088] This invention employs a hybrid loss function for deep supervised training, fusing cross-entropy and cross-union ratio (CIU) losses to more effectively locate and segment salient targets. To further explore the correlations between global pixels, an IOU loss function is introduced to calculate the overall structural similarity between the predicted results and the ground truth labels.

[0089] For example, the hybrid loss function is a weighted sum of the cross-entropy loss function and the IOU loss function:

[0090] L=λ1LBCE (P, G) + λ1L IOU (P, G)

[0091] Where λ1 and λ2 are the weight coefficients of the cross-entropy loss function and the IOU loss function, respectively, both set to 1, L BCE (P, G) represents the cross-entropy loss function of the saliency map, L IOU (P, G) represents the saliency map IOU loss function, where P is the predicted result of the saliency map and G is the ground truth label.

[0092] The cross-entropy loss function is widely used in classification problems, and its definition is as follows:

[0093]

[0094] Where l represents the l-th pixel in the image, and n represents the total number of pixels in the image.

[0095]

[0096] Using the aforementioned hybrid loss function as the optimization objective, and through deep training, the loss function is minimized to optimize the network model.

[0097] This invention also provides a salient target detection device based on a multi-scale dilated convolutional neural network, such as... Figure 5 As shown, the device includes:

[0098] Multi-scale encoding module: configured to take the RGB image to be processed as the input image, input the input image into a pre-trained ResNet34 network model, and extract the multi-scale features of the input image;

[0099] Dilated Residual Convolution Module: Configured to input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input to the second convolutional layer, each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module;

[0100] The salient feature acquisition module is configured to input the fused features into multiple channel attention modules respectively to obtain multiple salient features;

[0101] The saliency map generation module is configured to perform dimensionality reduction activation on each saliency feature to generate a saliency map, and uses a hybrid loss function that combines cross-entropy and cross-union loss for deep supervised training.

[0102] The specific embodiments described above only illustrate the design principles of the present invention. The shapes and names of the components in this description may differ and are not limited. Therefore, those skilled in the art can modify or make equivalent substitutions to the technical solutions described in the foregoing embodiments; and these modifications and substitutions do not depart from the inventive spirit and technical solutions of the present invention, and should all fall within the protection scope of the present invention.

Claims

1. A salient target detection method based on a multi-scale dilated convolutional neural network, characterized in that, Includes the following steps: Step S1: Use the RGB image to be processed as the input image, input the input image into the pre-trained ResNet34 network model, and extract the multi-scale features of the input image; Step S2: Input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input into the second convolutional layer, and each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module; Step S3: Input the fused features into multiple channel attention modules respectively to obtain multiple saliency features; Step S4: Dimensionally reduce and activate each salient feature to generate a saliency map, and perform deep supervised training using a hybrid loss function that combines cross-entropy and cross-union loss; In step S2, the following is stated: Both the first and second dilated convolution modules perform batch normalization and ReLU nonlinear activation operations after dilation and convolution; the fill rate of the first dilated convolution module is... The fill rate of the second dilated convolution module is ; The output of the branch is F. i , Where i represents the number of branches containing residual modules and dilated convolutions, and j represents the number of multi-scale feature branches. This represents multi-scale dimensionality reduction features. This indicates the convolution operation performed by the residual module. , These represent the fill rate respectively. , The dilated convolution operation, where ∑ represents the pixel-level additive fusion operation; The output F of each dilated convolution branch i Inputting the second convolutional layer yields fused features that include contextual information from the multi-scale features. ,in, denoted as dimensionality reduction convolution operation, and C represents channel-dimensional feature fusion operation.

2. The method as described in claim 1, characterized in that, The channel attention module includes an average pooling module, a max pooling module, a first convolutional branch, a second convolutional branch, and a third convolutional module. The fusion feature is input into the average pooling module and the max pooling module respectively. The results processed by the average pooling module and the max pooling module are fused to obtain a first fusion feature. The first fusion feature is input into the first convolutional branch and the second convolutional branch respectively. The first fusion feature, the processing result obtained from the first convolutional branch, and the processing result obtained from the second convolutional branch are fused to obtain a second fusion feature. The second fusion feature is input into the third convolutional module to obtain a saliency feature.

3. The method as described in claim 2, characterized in that, In step S3, the following is stated: The first convolutional branch and the second convolutional branch have different kernel sizes; the first convolutional branch includes a 3×3 convolutional layer, a fully connected layer, and a sigmoid layer. The first fused feature is input into the first convolutional branch to obtain the first degree of influence of different channel levels on the saliency information. The second convolutional branch includes a 5×5 convolutional layer, a fully connected layer, and a sigmoid layer. The first fused feature is input into the second convolutional branch to obtain the second degree of influence of different channel levels on the saliency information. ; to the first degree of impact Second degree of influence First fusion feature The integration will be carried out in the following manner: in, This indicates that the coefficients are multiplied. This represents the saliency feature of the channel attention module output of the multi-scale branch j.

4. The method as described in claim 3, characterized in that, In step S4, the hybrid loss function L is in, These are the cross-entropy loss functions. The weight coefficients of the IOU loss function, The cross-entropy loss function represents the saliency plot. The significance graph IOU loss function is represented. These are the prediction results from the saliency plot. These are the truth labels of the saliency plot; in, Indicates the first element in the input image. 1 pixel, This indicates that the input image has a total of n pixels. Indicates the first The ground truth labels of the saliency map for each pixel. Indicates the first Prediction results of the saliency map of pixels.

5. A salient target detection device based on a multi-scale dilated convolutional neural network, characterized in that, include: Multi-scale encoding module: configured to take the RGB image to be processed as the input image, input the input image into a pre-trained ResNet34 network model, and extract the multi-scale features of the input image; Dilated Residual Convolution Module: Configured to input the multi-scale features into the dilated residual convolution module to obtain fused features including contextual information of the multi-scale features; the dilated residual convolution module includes a first convolutional layer, multiple branches connected to the first convolutional layer, and a second convolutional layer; wherein the output of each branch is input to the second convolutional layer, each branch includes a residual module connected to the first convolutional layer, a first dilated convolutional module connected to the residual module, and a second dilated convolutional module, and the output of the branch is obtained based on the output of the first dilated convolutional module and the output of the second dilated convolutional module; The salient feature acquisition module is configured to input the fused features into multiple channel attention modules respectively to obtain multiple salient features; Saliency map generation module: configured to reduce the dimensionality of each saliency feature and generate a saliency map, using a hybrid loss function that combines cross-entropy and cross-union loss for deep supervised training; The dilated residual convolution module, wherein: Both the first and second dilated convolution modules perform batch normalization and ReLU nonlinear activation operations after dilation and convolution; the fill rate of the first dilated convolution module is... The fill rate of the second dilated convolution module is ; The output of the branch is F. i , Where i represents the number of branches containing residual modules and dilated convolutions, and j represents the number of multi-scale feature branches. This represents multi-scale dimensionality reduction features. This indicates the convolution operation performed by the residual module. , These represent the fill rate respectively. , The dilated convolution operation, where ∑ represents the pixel-level additive fusion operation; The output F of each dilated convolution branch i Inputting the second convolutional layer yields fused features that include contextual information from the multi-scale features. ,in, denoted as dimensionality reduction convolution operation, and C represents channel-dimensional feature fusion operation.

6. A computer-readable storage medium storing a plurality of instructions; the plurality of instructions being loaded by a processor and executing the method as claimed in any one of claims 1-4.

7. An electronic device, characterized in that, The electronic device includes: A processor is used to execute multiple instructions; Memory, used to store multiple instructions; The plurality of instructions are to be stored in the memory and loaded by the processor and executed as described in any one of claims 1-4.

Images