A low-altitude remote sensing image real-time semantic segmentation method

Abstract

The application discloses a kind of real-time semantic segmentation methods for low-altitude remote sensing image, the convolution unit in detail branch is optimized with LegoNet convolution unit, while semantic branch adopts the lightweight residual structure network of Stem module, lightweight residual module and triple attention module containing LegoNet convolution unit optimization, to construct improved BiSeNet V2 network model, then with improved BiSeNet V2 network model as segmentation network model, extract the detail features and semantic features of low-altitude remote sensing image, through guide aggregation layer to the weighted fusion of detail features and semantic features, while using bilinear interpolation to restore to image original resolution size, and based on adaptive threshold loss function for each pixel classification, finally complete segmentation. The semantic segmentation method of the application can effectively identify the building area in remote sensing image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a real-time semantic segmentation method for low-altitude remote sensing images. Background Technology

[0002] With the continuous development of aerospace technology, high-resolution remote sensing images can clearly reflect the geometric structure and rich spatial information of objects. Remote sensing image segmentation technology plays an important role in fields such as intelligence reconnaissance, agricultural production, topographic mapping, and daily life. Semantic segmentation of remote sensing images relies on shallow features such as image location and spectrum, as well as abstract semantic features, to segment the region to be identified and obtain the semantic ground truth value of each pixel. Traditional semantic segmentation methods often have great limitations in practical applications due to the large number of parameters. While lightweight semantic segmentation networks can improve real-time inference speed, they often sacrifice accuracy for efficiency, resulting in poor performance in low-altitude remote sensing image segmentation.

[0003] Remote sensing images suffer from difficulties in object identification due to the phenomena of "same object, different spectrum" and "same spectrum, different object". Furthermore, the differences between different land cover types are small due to mutual occlusion between shadows and objects. In addition, remote sensing targets themselves are characterized by small size, multiple scales, and dense distribution. Operations such as pooling in the feature extraction process will reduce image resolution when expanding the receptive field, and are prone to losing target feature information, so the segmentation effect is not ideal. Although BiSeNet V2 can effectively extract the differential features of objects to a certain extent, it ignores the cross-latitude interaction of feature map channels and location feature information, resulting in relatively coarse segmentation results when applied to low-altitude remote sensing images. It has problems such as low edge segmentation and localization accuracy and small target segmentation accuracy. Summary of the Invention

[0004] This invention provides a real-time semantic segmentation method for low-altitude remote sensing images, which solves the technical problems of low edge localization accuracy and poor segmentation accuracy of small targets when the existing BiSeNet V2 algorithm is applied to low-altitude remote sensing image segmentation.

[0005] This invention can be achieved through the following technical solutions:

[0006] A real-time semantic segmentation method for low-altitude remote sensing images is proposed. The convolutional units in the detail branch are optimized using LegoNet convolutional units, while the semantic branch employs a lightweight residual network structure containing a Stem module optimized from LegoNet convolutional units, a lightweight residual module, and a triple attention module to construct an improved BiSeNet V2 network model. This improved BiSeNet V2 network model is then used as the segmentation network model to extract detail and semantic features from the low-altitude remote sensing images. A guided aggregation layer performs weighted fusion of detail and semantic features, while bilinear interpolation is used to restore the original image resolution. Finally, an adaptive threshold loss function is used to classify each pixel, completing the segmentation process.

[0007] Furthermore, the semantic branch sequentially includes a first unit and three second units with the same structure. The first unit sequentially includes a Stem module optimized by the LegoNet convolutional unit, two lightweight residual modules with a stride of 1, and a triple attention module. The first unit sequentially includes lightweight residual modules with a stride of 2 and a stride of 1, and a triple attention module.

[0008] Furthermore, the lightweight residual module with a stride of 2 first passes the original feature map through a LegoNet convolutional unit with a kernel size of 3×3 and a stride of 2, then through a ReLU non-linear activation function layer, followed by a LegoNet convolutional unit with a kernel size of 3×3 to obtain the output feature map. Then, the original feature map is passed through a traditional upscaling convolutional layer with a kernel size of 1×1 and a stride of 2 and then short-connected with the output feature map. Finally, it passes through a ReLU non-linear activation function layer to obtain the final feature map.

[0009] The lightweight residual module with a stride of 1 first passes the original feature map sequentially through a LegoNet convolutional layer with a kernel size of 3×3, a ReLU non-linear activation function layer, and another LegoNet convolutional layer with a kernel size of 3×3 to obtain the output feature map. Then, the original feature map and the output feature map are concatenated using a residual shortcut, and finally, the final feature map is obtained by passing it through a ReLU non-linear activation function layer.

[0010] Furthermore, by utilizing semantic branches, the input image with a resolution of 1024×512×3 is first passed through the Stem module optimized by the LegoNet convolutional unit, two lightweight residual modules with a stride of 1, and a triple attention module to generate a feature map with a resolution of 256×128×16.

[0011] Next, the 256×128×16 feature map is passed sequentially through lightweight residual modules with strides of 2 and 1, and a triple attention module to generate a 128×64×32 feature map; then, the 128×64×32 feature map is passed sequentially through lightweight residual modules with strides of 2 and 1, and a triple attention module to generate a 64×32×64 deep feature map; finally, the 64×32×64 feature map is passed sequentially through lightweight residual modules with strides of 2 and 1, and a triple attention module to generate a 32×16×128 semantic feature map.

[0012] Using the detail branch, the input image with a resolution of 1024×512×3 is first subjected to two LegoNet convolutions to generate a 512×256×32 feature map; then the 512×256×32 feature map is subjected to three LegoNet convolutions to generate a 256×128×64 feature map; finally, the 256×128×64 feature map is subjected to three LegoNet convolutions to generate a 128×64×128 detail feature map.

[0013] Furthermore, the expression for the adaptive threshold loss function is:

[0014]

[0015] Where L represents the total loss function, l m Represents the main loss function, l i Let X represent the auxiliary loss function for the i-th stage. i This represents the feature map of the i-th stage; the expressions for the main loss function and the auxiliary loss function are:

[0016]

[0017] Where M represents the total number of pixels, C represents the total number of categories, and y i Let p represent the label value of the i-th pixel, ε{} represent the condition that the value is 1 if the condition is met and 0 otherwise. ij Let λ represent the probability that the i-th pixel is predicted to be of class j, and let λ represent the adaptive threshold.

[0018] The beneficial technical effects of this invention are as follows:

[0019] 1) A lightweight residual network containing LegoNet convolutional units and triple attention modules is used as the feature extraction network for semantic branches to preserve the detailed information of deep feature maps and enhance the model's ability to extract features from edge regions and small targets in remote sensing images.

[0020] 2) Optimize the detail branches using LegoNet convolutional units to further reduce the algorithm's weight while ensuring the model's feature extraction capabilities, thus meeting the requirements for real-time segmentation of high-resolution aerial images;

[0021] 3) Adaptive threshold loss function is used to supervise network training, further forcing the network to pay more attention to small targets with fewer pixels and independent regions. Attached Figure Description

[0022] Figure 1 This is a diagram of the overall network structure of the present invention;

[0023] Figure 2 This is a schematic diagram of the LegoNet convolutional unit of the present invention;

[0024] Figure 3 This is a schematic diagram of the Stem module structure of the present invention;

[0025] Figure 4 This is a structural diagram of the lightweight residual module of the present invention;

[0026] Figure 5 This is a schematic diagram of the triple attention module structure of the present invention;

[0027] Figure 6 The results of land structure segmentation using the semantic segmentation method of this invention are shown. Detailed Implementation

[0028] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, this invention provides a real-time semantic segmentation method for low-altitude remote sensing images. The convolutional units in the detail branch are optimized using LegoNet convolutional units, while the semantic branch employs a lightweight residual network structure containing a Stem module optimized from LegoNet convolutional units, a lightweight residual module, and a triple attention module to construct an improved BiSeNet V2 network model. This improved BiSeNet V2 network model is then used as the segmentation network model to extract detail and semantic features from the low-altitude remote sensing image. A guided aggregation layer performs weighted fusion of detail and semantic features, while bilinear interpolation is used to restore the image to its original resolution. Finally, an adaptive threshold loss function is used to classify each pixel, completing the segmentation process.

[0030] Specifically as follows:

[0031] Step 1: Construct a segmentation network model

[0032] S1: Replace the traditional convolutional layers in the detail branch with LegoNet convolutional units to reduce computational costs while fully extracting shallow spatial information from low-altitude remote sensing images.

[0033] The LegoNet convolutional unit mainly consists of three steps: splitting, transformation, and merging. In the splitting stage, the feature map is split into a set of multiple feature sub-maps. In the transformation stage, all feature sub-maps are convolved with each Lego convolutional kernel to obtain multiple Lego feature sub-maps. In the merging stage, the multiple Lego feature sub-maps are concatenated and merged to obtain the output feature map. Therefore, the detail branch optimized by the LegoNet convolutional unit can fully extract shallow spatial information of the image while reducing the computational cost of the model.

[0034] Specifically, the input image with a resolution of 1024×512×3 is first subjected to two LegoNet convolutions to generate a 512×256×32 feature map; then, the 512×256×32 feature map is subjected to three LegoNet convolutions to generate a 256×128×64 feature map; finally, the 256×128×64 feature map is subjected to three LegoNet convolutions to generate a 128×64×128 feature map; all of the above LegoNet convolutions are optimized convolution calculations of LegoNet convolutional units;

[0035] S2: The Stem module is optimized with LegoNet convolutional units, and a triple attention module is added after the lightweight residual module. A lightweight residual structure network containing LegoNet convolutional units and a triple attention module is constructed as a feature extraction network for semantic branches, which enhances the model's ability to extract features from edge regions and small targets in remote sensing images.

[0036] like Figure 3 As shown, the Stem module optimized using LegoNet convolutional units consists of two downsampling processes and an up-dimensional convolution part. LegoNet convolution is used to optimize the downsampling and up-dimensional convolution operations in turn, thereby reducing the computational cost.

[0037] like Figure 4 As shown, the lightweight residual module includes two lightweight residual structures with stride of 2 and stride of 1. The lightweight residual module with stride of 2 first passes the original feature map through a LegoNet convolutional unit with a kernel size of 3×3 and a stride of 2, then through a ReLU non-linear activation function layer, and then through a LegoNet convolutional layer with a kernel size of 3×3 to obtain the output feature map. Then, the original feature map passes through a traditional upscaling convolutional layer with a kernel size of 1×1 and a stride of 2 and is then short-connected with the output feature map. Finally, it passes through a ReLU non-linear activation function layer to obtain the final feature map.

[0038] The lightweight residual module with a stride of 1 first passes the original feature map sequentially through a LegoNet convolutional layer with a kernel size of 3×3, a ReLU non-linear activation function layer, and another LegoNet convolutional layer with a kernel size of 3×3 to obtain the output feature map. Then, the original feature map and the output feature map are shortcut concatenated, and finally, the final feature map is obtained by passing it through a ReLU non-linear activation function layer.

[0039] like Figure 5 As shown, the triple attention module achieves cross-dimensional interaction through three branches, capturing the relationship between the three tensors (C, H), (C, W), and (H, W) through rotation respectively;

[0040] The first branch is used to construct the relationship between the spatial dimension H and the channel dimension C: First, the H-axis of the input feature map is swapped with the C-axis. Then, the tensor of the C-dimensional dimension is reduced to two dimensions by max pooling and average pooling, respectively. Then, H×C interactive attention weights are generated by 1×1 convolution and Sigmoid activation function and multiplied with the input feature map. Finally, the H-axis of the output tensor is swapped with the C-axis to keep it consistent with the input.

[0041] The second branch is used to construct the relationship between the spatial dimension W and the channel dimension C: First, the W-axis of the input feature map is swapped with the C-axis. Then, the tensor of the C-dimensional dimension is reduced to two dimensions by max pooling and average pooling, respectively. Then, W×C interactive attention weights are generated by 1×1 convolution and Sigmoid activation function and multiplied with the input feature map. Finally, the W-axis of the output tensor is swapped with the C-axis to keep it consistent with the input.

[0042] The third branch is used to construct the relationship between spatial dimensions W and H: the input tensor is directly reduced from C-dimensional tensor to two dimensions through max pooling and average pooling, and then W×H interactive attention weights are generated through 1×1 convolution and sigmoid activation function and multiplied with the input feature map; finally, the output feature map is obtained by summing and averaging the outputs of the three branches; the triple attention module can well preserve spatial and channel information, extract semantic relationships in different dimensions, realize cross-dimensional interaction, and further enhance the network's ability to extract deep semantic features.

[0043] Specifically, the input image with a resolution of 1024×512×3 is first processed through a Stem module optimized by a LegoNet convolutional unit and two lightweight residual modules with a stride of 1. A triple attention module is then added to this layer to re-extract semantic relationship information in different dimensions, generating a feature map with a resolution of 256×128×16. Next, the 256×128×16 feature map is processed sequentially through lightweight residual modules with strides of 2 and 1, and the triple attention module to generate a 128×64×32 feature map. Then, the 128×64×32 feature map is processed sequentially through lightweight residual modules with strides of 2 and 1, and the triple attention module to generate a 64×32×64 deep feature map. Finally, the 64×32×64 feature map is processed sequentially through lightweight residual modules with strides of 2 and 1, and the triple attention module to generate a 32×16×128 semantic feature map.

[0044] S3: Classification and Recognition

[0045] The detail feature map output from the detail branch and the semantic feature map output from the semantic branch are weighted and fused through a guided aggregation layer. Simultaneously, bilinear interpolation is used to restore the image to its original resolution. Furthermore, each pixel is classified based on an adaptive threshold loss function, completing the segmentation process. The specific steps are as follows:

[0046] S31: The 128×64×128 detail feature map output from the detail branch and the 32×16×128 semantic feature map output from the semantic branch are respectively processed by 3×3 depthwise separable convolution and 1×1 convolution to obtain detail feature map and semantic feature map;

[0047] Meanwhile, the semantic branch generates semantic weights through 3×3 convolution and 4x bilinear interpolation upsampling, and multiplies them with the detail feature map to obtain a new 128×64×128 detail feature map. The detail branch generates detail weights through 3×3 convolution and average pooling, and multiplies them with the semantic feature map to obtain a new 32×16×128 semantic feature map.

[0048] Then, the 32×16×128 semantic feature map is upsampled by 4 times bilinear interpolation to generate a 128×64×128 semantic feature map, which is added to the 128×64×128 detail feature map and then convolved by 3×3 to generate a 128×64×128 fused feature map.

[0049] Finally, the 128×64×128 fused feature map is used to generate a 1024×512×N semantic segmentation prediction map by passing the traditional 3×3 convolution and 8x bilinear interpolation upsampling method, where N is the number of categories of the predicted pixels.

[0050] S32: An adaptive threshold loss function is used to classify each pixel. The training weights of high-loss pixel samples are increased to improve the model's attention to small, independent target objects with a low pixel count. Specifically, this is achieved by calculating the semantic category prediction probability p. ij The loss is calculated by taking the mean loss of pixel samples smaller than a threshold λ and then backpropagating it, thus solving the problem of uneven sample distribution. The expression for the loss function is:

[0051]

[0052] Where L represents the total loss function, l m Represents the main loss function, l i Let X represent the auxiliary loss function for the i-th stage. i This represents the feature map of the i-th stage; the expressions for the main loss function and the auxiliary loss function are:

[0053]

[0054] Where M represents the total number of pixels, C represents the total number of categories, and y i Let p represent the label value of the i-th pixel, ε{} represent the condition that the value is 1 if the condition is met and 0 otherwise. ij λ represents the probability value that the i-th pixel is predicted to be of class j, and λ represents the adaptive threshold, which takes the value λ∈(0,1).

[0055] Step two involves training and validating the constructed segmentation network model, and then using the trained segmentation network model to perform real-time semantic segmentation of low-altitude remote sensing images.

[0056] The following section utilizes the real-time semantic segmentation method of the present invention to perform semantic segmentation on low-altitude remote sensing images.

[0057] Example 1

[0058] The experimental materials and equipment, parameter settings, and evaluation indicators for this embodiment are as follows:

[0059] 1) Experimental materials and equipment

[0060] The UAVid 2020 dataset was used to train and validate the algorithm in the experiment. The dataset contains 420 high-resolution images of drones, with 200 images for training, 70 for validation, and 150 for testing. The images cover eight object categories, including vegetation, buildings, cars, and pedestrians. During training, the dataset was preprocessed, and data augmentation techniques such as random cropping, random scaling, and horizontal flipping were applied to the images to improve the model's generalization ability.

[0061] The experiment was implemented using the deep learning framework PyTorch. The experimental platform was configured as follows: an Intel Core i5-11500 processor with 16GB of RAM; an Nvidia RTX 3060 dedicated graphics card with 12GB of VRAM; CUDA 10.2 and CUDA NumPy 7.51 to accelerate GPU computation; Python packages such as NumPy for data processing; and Matplotlib library for visualization of segmentation results.

[0062] 2) Parameter settings

[0063] The model training process employs an adaptive threshold loss function to supervise network training. During training, difficult pixel samples with diversity and high loss are automatically selected as training samples. The selection method for difficult pixel samples is as follows: after sorting by loss value in descending order, if the number of samples greater than the threshold exceeds the initial set value, all samples greater than the threshold are selected; otherwise, samples are selected sequentially according to the set value. The adaptive threshold is the semantic category prediction probability value of the 512×batchsize difficult pixel sample, where batchsize represents the number of training image samples in a single batch. The formula for setting the sample number is shown below, where size_0 and size_1 represent the width and height of the input image, respectively:

[0064]

[0065] The model training parameters were set as follows: batch size was set to 6; number of training epochs was set to 120; the initial learning rate was set to 0.01, and the learning rate was reduced to 0.001 after 50 epochs; the Adam optimizer was used to update the network training weights.

[0066] 3) Evaluation Indicators

[0067] When lightweight semantic segmentation algorithms are applied to image segmentation in complex traffic scenes, both algorithm accuracy and operational efficiency should be considered in light of the specific circumstances. Metrics for measuring the accuracy of semantic segmentation algorithms include pixel accuracy (PA), mean pixel accuracy (mPA), and mean intersection over union (mIoU). While mPA and mIoU are performance metrics based on the global image level, mIoU better reflects the regional consistency of objects in a traffic scene image and more accurately represents the actual segmentation performance. Therefore, mIoU is chosen as the performance metric for evaluating algorithm accuracy, and its calculation formula is shown below:

[0068]

[0069] In addition, the actual running efficiency of the algorithm is represented by the real-time inference speed metric FPS (Frames Per Second), and its calculation formula is shown in the following formula:

[0070]

[0071] In the formula: k represents the number of pixel categories; t i The category represents the total number of pixels i; p ii p represents the number of pixels with the actual category i and the predicted category i; ji T represents the number of pixels whose actual category is j and predicted category is i; N represents the number of images; T represents the number of pixels whose actual category is j and predicted category is i. j This represents the time required for the algorithm to process the j-th image.

[0072] 4) Result Evaluation

[0073] Table 1 shows the performance comparison between BiSeNet V2 and the improved method. As can be seen from Table 1, the improved algorithm has improved the IoU of each category to varying degrees. Specifically, the IoU of small objects such as stationary cars, moving cars, and pedestrians has increased by 4.1%, 4.5%, and 4.4%, respectively, and the mIoU has increased by about 3.3%. With an input image resolution of 1024 pixels × 512 pixels, the inference speed reaches 72.8 FPS. Although the inference speed is lower than that of the algorithm before the improvement, it still meets the requirements of real-time image segmentation.

[0074] Table 1 Performance comparison between BiSeNet V2 algorithm and the algorithm of this invention

[0075]

[0076] Example 2

[0077] The algorithm uses data from the Earth Surface Building Semantic Segmentation Dataset from the Alibaba Cloud Tianchi Algorithm Competition. 20,000 images with a resolution of 512×512 pixels were randomly selected from this dataset for experiments. The dataset was divided into three parts: a training set of 15,000 images, a validation set of 4,000 images, and a test set of 1,000 images. The evaluation metrics were the same as in Example 1. The model parameters were set as follows: batch size of 16; 80 epochs; initial learning rate of 0.001, which was reduced to 0.0005 after 40 epochs; and the AdamW optimizer was used to update the network training weights with a weight decay coefficient of 0.01.

[0078] Figure 6The images show the results of building segmentation using BiSeNet V2 and an improved BiSeNet V2 algorithm. (a) is the original remote sensing image, (b) is a manually labeled image, (c) is the segmentation result using the BiSeNet V2 algorithm, and (d) is the segmentation result of the same image using the algorithm proposed in this invention. The results show that the BiSeNet V2 segmentation suffers from lost segmentation edges and over-segmentation. In contrast, the algorithm proposed in this invention can effectively identify building regions in remote sensing images, with clear object boundaries, fewer segmentation errors, and good results even when segmenting buildings at multiple different scales.

[0079] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time semantic segmentation method for low-altitude remote sensing images, characterized in that: The convolutional units in the detail branch are optimized using LegoNet convolutional units, while the semantic branch adopts a lightweight residual network structure containing the Stem module optimized from LegoNet convolutional units, a lightweight residual module, and a triple attention module to construct an improved BiSeNet V2 network model. Then, the improved BiSeNet V2 network model is used as the segmentation network model to extract detail features and semantic features from low-altitude remote sensing images. The detail features and semantic features are weighted and fused through a guided aggregation layer. At the same time, bilinear interpolation is used to restore the original resolution of the image, and each pixel is classified based on an adaptive threshold loss function to complete the segmentation. The semantic branch includes a first unit and three second units with the same structure. The first unit includes a Stem module optimized by the LegoNet convolutional unit, two lightweight residual modules with a stride of 1, and a triple attention module. The first unit also includes lightweight residual modules with a stride of 2 and a stride of 1, and a triple attention module. Using the detail branch, the input image with a resolution of 1024×512×3 is first subjected to two LegoNet convolutions to generate a 512×256×32 feature map; then the 512×256×32 feature map is subjected to three LegoNet convolutions to generate a 256×128×64 feature map; finally, the 256×128×64 feature map is subjected to three LegoNet convolutions to generate a 128×64×128 detail feature map.

2. The real-time semantic segmentation method for low-altitude remote sensing images according to claim 1, characterized in that: The lightweight residual module with a stride of 2 first processes the original feature map through a convolution kernel of size 3.

3. A LegoNet convolutional unit with a stride of 2, followed by a ReLU non-linear activation function layer, and then a convolutional kernel with a size of 3. The output feature map is obtained by using 3 LegoNet convolutional units, and then the original feature map is processed by convolutional kernels with a size of 1.

1. After a traditional dimensionality-upgrading convolutional layer with a stride of 2, a shortcut connection is made with the output feature map, and finally, the final feature map is obtained by passing through a ReLU non-linear activation function layer. The lightweight residual module with a stride of 1 first passes the original feature map sequentially through convolutional kernels of size 3. The LegoNet convolutional layers, ReLU non-linear activation function layers, and convolutional kernels are all 3 in size. The output feature map is obtained by the 3rd convolutional layer of LegoNet. Then, the original feature map and the output feature map are connected by residual shortcut. Finally, the final feature map is obtained by passing through the ReLU non-linear activation function layer.

3. The real-time semantic segmentation method for low-altitude remote sensing images according to claim 1, characterized in that: Using semantic branching, first, the resolution is 1024. 512 The input image (3) is sequentially processed through a LegoNet convolutional unit-optimized Stem module, two lightweight residual modules with a stride of 1, and a triple attention module to generate a resolution of 256. 128 16 feature maps; Next, 256 128 The 16 feature maps are sequentially processed by lightweight residual modules with strides of 2 and 1, and then by a triple attention module to generate 128. 64 32 feature maps; then 128 64 The 32 feature maps are sequentially processed by lightweight residual modules with strides of 2 and 1, and then by a triple attention module to generate 64. 32 64 deep feature maps; then 64 32 The 64 feature maps are sequentially processed by lightweight residual modules with strides of 2 and 1, and then by a triple attention module to generate 32. 16 Semantic feature map of 128.

4. The real-time semantic segmentation method for low-altitude remote sensing images according to claim 1, characterized in that: The expression for the adaptive threshold loss function is: in, Represents the total loss function. Represents the main loss function. Indicates the first Each stage's auxiliary loss function Indicates the first Feature maps for each stage; expressions for the main loss function and auxiliary loss function are: in This indicates the total number of pixels. Indicates the total number of categories. Indicates the first The label value of each pixel. This means that the value is 1 if the condition is met, and 0 otherwise. Indicates the first The predicted pixel is the [number]th pixel. The probability value of the class. This indicates an adaptive threshold.

Images