Semantic segmentation method and device based on double-branch feature fusion
By employing a dual-branch feature fusion method, low-level details and high-level semantic information are extracted separately. The final segmentation representation is generated through the fusion module FFM and the attention mechanism, which solves the problems of high computational cost and insufficient information fusion in existing technologies, thereby improving the accuracy and robustness of semantic segmentation.
Patent Information
- Application Number
- CN202211623747.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-16
- Publication Date
- 2026-02-24
- Estimated Expiration
- 2042-12-16
AI Technical Summary
Existing semantic segmentation networks are computationally expensive when processing high-resolution images, tend to ignore the detailed appearance of boundaries and small objects, and have difficulty effectively integrating low-resolution and high-resolution information.
A dual-branch feature fusion method is adopted, which extracts low-level detail information and high-level semantic information through detail branch and semantic branch respectively. Then, the fusion module FFM introduces semantic information into the low-level features and detail information into the high-level features. Combined with spatial and channel attention mechanisms, the final segmentation representation is generated.
It improves the accuracy and robustness of semantic segmentation, reduces computational costs, effectively integrates low-level details and high-level semantic information, and enhances feature representation capabilities.
Smart Images
Figure CN116229056B_ABST
Abstract
Description
[0001] This invention relates to the field of computer vision technology, and in particular to semantic segmentation methods, apparatus, and devices based on bi-branch feature fusion. Background Technology
[0002] Semantic segmentation is a key task in computer vision, aiming to assign dense labels to all pixels in an image—a process of moving from concrete to abstract. In recent years, convolutional neural network (CNN) architectures have seen continuous innovation, achieving impressive results. Fully convolutional networks (LCNs) have demonstrated that end-to-end, pixel-to-pixel CNNs surpass previous state-of-the-art techniques. LCNs involve transforming fully connected layers into convolutional layers and upsampling through deconvolution; skip connections allow semantic information to be combined with representational information, resulting in accurate and fine-grained segmentation. However, as network depth increases, the receptive field of fully convolution grows slowly, and this limited receptive field cannot fully simulate the long-distance relationships between pixels in an image. UNet combines low-resolution information (object category) and high-resolution information (precise segmentation and localization), making it perfectly suited for medical image segmentation. While convolution performs well in many recognition tasks, it has not been widely applied to specific tasks due to limited training scale and uncertain network size, thus hindering its development. PSANet introduces an attention mechanism in the decoder, using adaptive attention to connect pixels at each location in the feature map, promoting information transfer and improving segmentation performance in complex scenes. DeepLabV3+ uses a hollow spatial pyramid pooling module to extract multi-scale features of objects, explicitly preserving high-resolution representations. However, this network only fuses one layer. Bisenet proposes that spatial paths can encode rich spatial and detail information, and utilizes a feature fusion module to fuse spatial and detail information. It also presents a new approach, emphasizing the need to consider spatial information while improving speed. This design also represents a rethinking of semantic segmentation backbone networks. This method can be applied not only to real-time semantic segmentation algorithms but also to other fields, especially when both spatial detail and contextual information are required. However, these methods often involve expensive computations. Furthermore, they typically use raw high-resolution images as input, further increasing computational cost, and previous network architectures easily overlook boundaries and detailed appearances around small objects, thus missing high-resolution information. Summary of the Invention
[0003] To address the shortcomings of existing technologies, this invention proposes a dual-channel network consisting of two branches: a detail branch and a semantic branch. This allows the network to separately acquire low-level detail information and high-level semantic information. Notably, the two branches in our network are not executed independently. We obtain detailed edge features by extracting texture information from the input image. The detail branch complements the semantic branch; the detail map is added to the semantic feature map to supplement the detail features. Finally, the two optimized features are merged into the final segmentation representation.
[0004] Includes the following steps;
[0005] S1. Construct the network framework, which consists of two parts: detail branches and semantic branches. Input the given image into the backbone network to extract semantic features. First, we reduce the size of the input image by 16 times through the encoder.
[0006] S2. Features extracted from the backbone network are passed through the dilated spatial convolution pooling pyramid. Its core idea is to aggregate receptive fields of different scales. The dilated spatial convolution pooling pyramid is also proposed to solve the problem of different scales for different segmentation targets. It consists of a 1×1 convolution kernel and three 3×3 dilated convolutions with dilation rates of 3, 6 and 12 respectively.
[0007] S3. Use a 1×1 convolution to reduce the number of channels, followed by a BN, ReLU activation function and Dropout; use bilinear interpolation to upsample it by 4 times.
[0008] S4. The detail branch extracts texture information from the input image to obtain detail edge features. The purpose is to extract spatial detail information and then use the edge features to enhance semantic information. As a supplement to the semantic branch, the detail branch will be added to the semantic feature map to supplement the detail features.
[0009] S5. The Fusion Module (FFM) is used to fuse high-level semantic and detailed information. The Fusion Module (FFM) is as follows: Figure 3 As shown, semantic information is introduced into the low-level features and detailed information is introduced into the high-level features, making subsequent fusion more effective and enhancing feature representation;
[0010] S6. Insert a detail head into the fused feature map to generate binary detail labels. Then, use the binary detail labels as a guide for the detail feature map to guide the underlying learning of spatial detail features. Finally, reduce the number of channels by performing 1×1 convolutions on the feature map, followed by a BN and ReLU activation function.
[0011] S7, concatenates the feature maps of S3 and S5; then inputs them into the decoder, the decoder as follows: Figure 1As shown in (b), the final result is obtained by performing 3×3 convolution, BN, ReLU, Dropout, and upsampling by 4 times.
[0012] S8 uses Binary cross entropy and Dice Loss, Focal Loss, and CE Loss to jointly optimize detailed learning.
[0013] This invention provides a semantic segmentation method, apparatus, and device based on dual-branch feature fusion. Compared with existing technologies, it has the following advantages:
[0014] Our actual network structure can obtain low-level detail information and high-level semantic information separately. We obtain detailed edge features by extracting texture information from the input image. The detail branch complements the semantic branch, obtaining detail information from the low-order network and semantic information from the high-order network, and then fusing them to avoid missing any type of information. High-order semantic information optimizes low-order edge information, and the two optimized features are then merged into the final segmentation representation. Secondly, we propose a fusion module (FFM) to fuse high-level semantic and detail information to enhance feature representation. Furthermore, the fusion module employs an attention mechanism to process feature maps from both branches to establish contextual dependencies in spatial and channel dimensions, which helps the network focus on more meaningful features. The feature maps extracted by the detail extraction branch generate the final prediction through the detail segmentation head to improve performance at a negligible cost. Simultaneously, a joint loss of Binarycross Entropy and Dice loss is used to guide shallow information to encode spatial information, performing feedback iterative optimization of the model's loss to minimize the final loss, thereby improving feature accuracy and robustness. Attached Figure Description
[0015] 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is the overall network structure diagram of the semantic segmentation method model based on dual-branch feature fusion in this invention patent, which includes the decoder and detail head.
[0017] Figure 2 This is a structural diagram of the spatial and channel attention mechanism in this invention patent.
[0018] Figure 3This is a structural diagram of the feature fusion module in this invention patent.
[0019] Figure 4 It is the device in this invention patent.
[0020] Figure 5 It is the device described in this invention patent. Detailed Implementation
[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.
[0022] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The semantic segmentation method, apparatus, and device for dual-branch feature fusion includes steps S1 to S8:
[0023] S1. Construct the network framework, which consists of two parts: detail branches and semantic branches. Input the given image into the backbone network to extract semantic features. First, we reduce the size of the input image by 16 times through the encoder.
[0024] S2. Features extracted from the backbone network are passed through the dilated spatial convolution pooling pyramid. Its core idea is to aggregate receptive fields of different scales. The dilated spatial convolution pooling pyramid is also proposed to solve the problem of different scales for different segmentation targets. It consists of a 1×1 convolution kernel and three 3×3 dilated convolutions with dilation rates of 3, 6 and 12 respectively.
[0025] S3. Use a 1×1 convolution to reduce the number of channels, followed by a BN, ReLU activation function and Dropout; use bilinear interpolation to upsample it by 4 times.
[0026] S4. The detail branch extracts texture information from the input image to obtain detail edge features. The purpose is to extract spatial detail information and then use the edge features to enhance semantic information. As a supplement to the semantic branch, the detail branch will be added to the semantic feature map to supplement the detail features.
[0027] S5, the Fusion Module (FFM) is used to fuse high-level semantic and detailed information, such as... Figure 3 As shown, semantic information is introduced into the low-level features and detailed information is introduced into the high-level features, making subsequent fusion more effective and enhancing feature representation;
[0028] S6. Insert detail heads into the fused feature map portion to generate binary classification detail labels. The detail heads are as follows: Figure 1As shown in (c), the binary classification detail labels are then used as a guide for the detail feature map to guide the underlying learning of spatial detail features; finally, the feature map part is sequentially processed by 1×1 convolution to reduce the number of channels, followed by a BN and ReLU activation function.
[0029] S7, concatenates the feature maps of S3 and S5; then inputs them into the decoder, the decoder as follows: Figure 1 As shown in (b), the final result is obtained by performing 3×3 convolution, BN, ReLU, Dropout, and upsampling by 4 times.
[0030] S8 uses Binary cross entropy and Dice Loss, Focal Loss, and CE Loss to jointly optimize detailed learning.
[0031] The following is a detailed description of each step.
[0032] In step S1, the network architecture is constructed, such as... Figure 1 As shown, a network framework is constructed, consisting of semantic branches and detail branches. The specific steps are as follows:
[0033] The backbone networks used in the feature extraction section are Xception, VGGNet, and ResNet18. Xception, VGGNet, and ResNet18 have appeared in many classic network architectures and have been widely recognized and proven. Furthermore, we need to construct a lightweight network, so we use Xception, VGGNet, and ResNet18 to extract features from images, thereby demonstrating the effectiveness of our proposed method.
[0034] In step S2, the features extracted from the backbone network are processed through ASPP. ASPP was proposed to address the issue of different scales for different segmentation targets. The specific steps are as follows:
[0035] The core idea of S201, the dilated spatial convolution pooling pyramid, is to concentrate the receptive field of multiple scales. The dilated spatial convolution pooling pyramid consists of one 1*1 convolution and three 3*3 dilated convolutions with dilation rates of 3, 6 and 12 respectively.
[0036] S202. Add the feature map obtained by the 1*1 convolution and the feature map obtained by the 3*3 dilated convolution with dilation rates of 3, 6 and 12 respectively. The final feature map is the output map of the dilated spatial convolution pooling pyramid.
[0037] In step S3, a 1×1 convolution is used to reduce the number of channels, followed by a BN, ReLU activation function and Dropout; bilinear interpolation is used to upsample it by a factor of 4.
[0038] In step S4, the detail branch extracts texture information from the input image to obtain detailed edge features. The purpose is to extract spatial detail information, and then use these edge features to enhance semantic information. As a supplement to the semantic branch, the detail branch is added to the semantic feature map to supplement the detail features. The specific steps are as follows:
[0039] S401. In our detail branch, the first step is to extract the texture information of the image. Learning robust texture representations against scale changes is crucial for texture recognition. In this patent, first-order gradient Sobel, Laplacian, and local binary mode methods are used.
[0040] S402. We fuse multiple convolutional layers used for texture representation and extract the complementary relationships between them by utilizing multi-texture information.
[0041] In step S5, the fusion module (FFM) is used to fuse high-level semantic and detailed information, such as... Figure 3 As shown, semantic information is introduced into the low-level features, and detailed information is introduced into the high-level features to make subsequent fusion more effective and enhance feature representation. The specific steps are as follows:
[0042] S501. Attention helps the model assign different weights to each part of the input, extracting more critical and important information, enabling the model to make more accurate judgments, without increasing the computational and storage overhead. For the shallow feature map S extracted by the backbone network, to further enrich its spatial details, we use a spatial attention mechanism to weight it to obtain S'. The spatial attention mechanism is as follows: Figure 2 As shown in (b), the calculation formula is as follows:
[0043]
[0044] S502. For the feature map D extracted from the detail branch, we use channel attention to calculate the attention weights, multiply them by D to obtain D', thereby enhancing the detail discrimination between different channels. The channel attention mechanism is as follows: Figure 2 As shown in (a), the calculation formula is as follows:
[0045]
[0046] S503, S', and D' have the same number of channels, and the shallow feature map and the detail feature map can be fused by adding them element by element. We add S' and D' to further enhance the detail information contained in the shallow feature map. The fused feature map is used as the input to the backbone network, while the feature map of the semantic branch is added to the detail feature map as the input to the semantic branch. The overall calculation process of the module can be represented by the following formula:
[0047] Si+1=S'i+D'i (3)
[0048] In step S6, a detail head is inserted into the fused feature map portion to generate binary classification detail labels. The detail head is as follows: Figure 1 As shown in (c), binary classification detail labels are then used as a guide for the detailed feature map, guiding the underlying learning of spatial detail features; finally, the feature map part is sequentially processed with 1×1 convolutions to reduce the number of channels, followed by a BN and ReLU activation function, as follows. Figure 1 (a) The specific steps are as follows:
[0049] S601. We first generate binary detail labels from the ground truth labels of semantic segmentation using Laplacian convolution. We then insert detail headers into the shallow feature layers to generate the binary detail labels, as shown in Figure 601. Figure 1 As shown in (c), binary classification detail labels are then used as guides for the detail feature map, guiding the lower-level learning of spatial detail features. Compared to the low-level feature results, the binary classification detail label map with detail guidance can encode more spatial details.
[0050] S602, we generate binary detail background labels from semantically segmented background labels using the detail aggregation module. This operation can be implemented using a 2D Laplacian convolution kernel and a 1×1 convolution. We use... Figure 1 The Laplacian convolution kernel shown generates detail feature maps with different strides to obtain multi-scale detail information. We then upsample the detail feature maps to their original size and fuse them with a trainable 1×1 convolution.
[0051] S603. Finally, we use a threshold of 0.1 to convert the predicted details into final binary detail labels with boundary and corner information. Since the number of detail pixels is much smaller than that of non-detail pixels, detail prediction is a classic balance problem. Because weighted cross-entropy always leads to coarse results, we use Binary cross-entropy and Dice loss to jointly optimize detail learning. Dice loss measures the overlap between the predicted map and the ground truth label. Furthermore, it is insensitive to the number of foreground / background pixels, meaning it can alleviate the class imbalance problem.
[0052] Therefore, for a predicted detail map with height H and width W, the detail loss L detail The formula is as follows:
[0053] L_ detail( p_ d g_ d) =L _bce (p_ d g_ d )+L_dice(p_d g_ d (4)
[0054] Where p_ d ∈R H×W Indicates prediction details and g_ d ∈R H×W L_ represents the corresponding real label. bce L_ represents the binary cross-entropy loss. dice The Dice loss is represented as follows:
[0055]
[0056] Where i represents the i-th pixel, and ∈ is the Laplacian smoothing term. We set ∈ = 1 to estimate the probability of a phenomenon that has not occurred before. For example... Figure 1 As shown, we use a detail head to generate detail feature maps, guiding shallow information to encode spatial information. The detail head includes a 3×3 Conv, BN, and ReLU, followed by a 1×1 convolution to obtain the output details. The detail head effectively enhances the feature representation. Finally, the learned detail features are fused with the contextual features of the deep blocks in the decoder for segmentation prediction. However, this branch is discarded during the inference stage. Therefore, this side information can easily improve the accuracy of the segmentation task without any inference cost.
[0057] In step S7, the feature maps from S3 and S5 are concatenated; then input into the decoder, which is as follows: Figure 1 As shown in (b), the final result is obtained by performing 3×3 convolution, BN, ReLU, Dropout, and upsampling by 4 times.
[0058] In step S8, Binary cross entropy, Dice Loss, Focal Loss, and CE Loss are used to jointly optimize the learning of details. Because the training process of the network model is a continuous optimization of the loss, the currently obtained loss is fed back into the network model for iterative optimization to reduce the loss and thus obtain more robust features.
[0059] This invention improves the accuracy and robustness of the model by refining the feature extraction method. Our actual network structure can obtain low-level detail information and high-level semantic information separately. Detail edge features are obtained by extracting texture information from the input image. The detail branch complements the semantic branch, obtaining detail information from the low-order network and semantic information from the high-order network, which are then fused to avoid missing any type of information. High-order semantic information optimizes low-order edge information, and the two optimized features are then merged into the final segmentation representation. Secondly, we propose a fusion module (FFM) to fuse high-level semantic and detail information to enhance the feature representation. Furthermore, the fusion module employs an attention mechanism to process feature maps from both branches, establishing spatial and channel-dimensional contextual dependencies, which helps the network focus on more meaningful features. The feature maps extracted by the detail extraction branch generate the final prediction through the detail segmentation head, improving performance at a negligible cost. This method constructs a novel and effective approach for semantic segmentation, providing a more efficient framework for semantic segmentation in practical applications.
[0060] The present invention also proposes an apparatus, such as Figure 4 It includes a semantic segmentation network training module that incorporates dual-branch feature fusion, and is also used to input the fused feature map into the decoder to finally obtain the sample prediction result.
[0061] The present invention also proposes a computer device, such as Figure 5 The device includes a processor, memory, network interface, display, and input device. When the processor executes the computer program, it implements the steps of the method described above. The processor provides computational and control capabilities. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a semantic segmentation method based on dual-branch feature fusion. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.
[0062] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.
Claims
1. Semantic segmentation methods, devices, and equipment based on dual-branch feature fusion. Its features include the following methods: Acquire high-resolution images and annotate them to obtain training, validation, and test samples; We construct a network framework consisting of detail branches and semantic branches. We input a given image into the backbone network to extract semantic features. First, we reduce the size of the input image by 16 times. Features extracted from the backbone network are passed through a dilated spatial convolution pooling pyramid, which consists of a 1×1 convolution kernel and three 3×3 dilated convolutions with dilation rates of 3, 6 and 12 respectively. We use a 1×1 convolution to reduce the number of channels, followed by a BN, ReLU activation function and Dropout; we use bilinear interpolation to upsample it by 4 times to obtain the feature map of the semantic branch. The detail branch extracts texture information from the input image to obtain detail edge features. The purpose is to extract spatial detail information and then use the edge features to enhance semantic information. As a supplement to the semantic branch, the detail branch is added to the semantic feature map to supplement the detail features. The fusion module (FFM) is used to fuse high-level semantic and detailed information. This paper introduces semantic information into low-level features and detailed information into high-level features to make subsequent fusion more effective and enhance feature representation. The fused feature map is inserted with a detail head to generate binary detail labels. These binary detail labels are then used as a guide for the detail feature map to guide the underlying learning of spatial detail features. Finally, the feature map is sequentially processed with 1×1 convolutions to reduce the number of channels, followed by a BN and ReLU activation function. The feature maps of the fusion module and the semantic branch are merged; then they are input into the decoder, which involves 3×3 convolution, BN, ReLU, and Dropout. Upsample by 4 times to obtain the final result; Binary cross entropy, Dice Loss, Focal Loss, and CE Loss are used to jointly optimize detailed learning; The network is trained to obtain a trained semantic segmentation algorithm model based on dual-branch feature fusion; the image to be tested is obtained and input into the trained segmentation model to obtain the prediction result of the image.
2. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion as described in claim 1, characterized in that, Create an image dataset, including: Acquire image samples; The collected image samples are labeled using image annotation tools. The result of semantic segmentation is to turn the image into color blocks with certain semantic information. Semantic segmentation technology can identify the semantic category of each color block and label each pixel with its corresponding label. A sample dataset is constructed using labeled image samples. The sample dataset is divided into a training dataset, a validation dataset, and a test dataset. The training dataset is then preprocessed.
3. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion according to claim 1, characterized in that, The backbone networks used in the feature extraction part are Xception, VGGNet, and ResNet18.
4. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion according to claim 1, characterized in that, The network consists of two parts: a semantic branch and a detail branch. The detail branch extracts texture information from the input image to obtain detailed edge features, with the aim of extracting spatial detail information. Then, the edge features are used to enhance the semantic information. As a supplement to the semantic branch, the detail branch is added to the semantic feature map to supplement the detail features.
5. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion according to claim 4, characterized in that, In our detail branch, we first extract the texture information of the image and learn robust texture representations, including first-order gradient Sobel, Laplacian, and local binary mode methods. We fuse multiple convolutional layers used for texture representation, utilize multi-texture information, extract the complementary relationships between them, and then use a fusion module to fuse semantic and detail information.
6. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion according to claim 5, characterized in that, The Fusion Module (FFM) is used to fuse high-level semantic and detail information. This paper introduces semantic information into low-level features and detail information into high-level features to make subsequent fusion more effective and enhance feature representation. For the shallow feature map S extracted by the backbone network, we use a spatial attention mechanism to weight it to obtain S' in order to further enrich its spatial details. For the feature map D extracted by the detail branch, we use channel attention to calculate the attention weights and multiply them with D to obtain D', which enhances the detail discrimination between different channels. S' and D' have the same number of channels. By adding them element by element, the fusion of shallow feature map and detail feature map can be achieved. The fused feature map part is inserted into the detail header to generate binary detail labels. Then, the binary detail labels are used as guides for the detail feature map to guide the lower layer to learn spatial detail features.
7. The semantic segmentation method, apparatus, and device based on dual-branch feature fusion according to claim 6, characterized in that, During the training of the network model, Binary cross entropy, Dice loss, and Focal loss are used to jointly optimize the learning of details. The training process of the network model is a continuous optimization of the loss, feeding the currently obtained loss back into the network model for continuous iterative optimization.
8. A semantic segmentation method, apparatus, or device based on dual-branch feature fusion, characterized in that, The device includes: Acquire high-resolution images and annotate them to obtain training, validation, and test samples; We construct a semantic segmentation network based on dual-branch feature fusion. The semantic segmentation network includes a backbone network, semantic branches, detail branches, and a fusion module. Given an image is input into the backbone network to extract semantic features. First, we reduce the size of the input image by 16 times. Features extracted from the backbone network are passed through a dilated spatial convolution pooling pyramid, which consists of a 1×1 convolution kernel and three 3×3 dilated convolutions with dilation rates of 3, 6 and 12 respectively. Use a 1×1 convolution to reduce the number of channels, followed by a BN, ReLU activation function and Dropout; use bilinear interpolation to upsample it by 4 times; The detail branch extracts texture information from the input image to obtain detail edge features. The purpose is to extract spatial detail information and then use the edge features to enhance semantic information. As a supplement to the semantic branch, the detail branch is added to the semantic feature map to supplement the detail features. The fusion module (FFM) is used to fuse high-level semantic and detailed information. This paper introduces semantic information into low-level features and detailed information into high-level features to make subsequent fusion more effective and enhance feature representation. The fused feature map is inserted with a detail head to generate binary detail labels. These binary detail labels are then used as a guide for the detail feature map to guide the underlying learning of spatial detail features. Finally, the feature map is sequentially processed with 1×1 convolutions to reduce the number of channels, followed by a BN and ReLU activation function. The feature maps of the fusion module and the semantic branch are merged; then they are input into the decoder, which involves 3×3 convolution, BN, ReLU, and Dropout. Upsample by 4 times to obtain the final result; Binary cross entropy, Dice Loss, Focal Loss, and CE Loss are used to jointly optimize detailed learning; The network is trained to obtain a trained semantic segmentation algorithm model based on dual-branch feature fusion; the image to be tested is obtained and input into the trained segmentation model to obtain the prediction result of the image.
9. A device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the steps of any one of claims 1 to 8 when executing a computer program.