Semantic segmentation method for high-resolution remote sensing images based on probabilistic graph representation of edges

Abstract

This invention discloses a high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation. The method includes: acquiring a remote sensing image set; augmenting the remote sensing image set; normalizing the remote sensing images in the augmented set; obtaining a training set from the normalized set; constructing a high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement; using this model to obtain semantic segmentation results for the remote sensing images; training the model using the training set; acquiring the remote sensing image to be segmented; inputting the image to be segmented into the trained model; and obtaining the semantic segmentation results. This method improves the recognition accuracy of small objects and complex boundary information in remote sensing images, achieving accurate semantic segmentation of remote sensing ground feature edges.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and in particular to a high-resolution remote sensing image semantic segmentation method based on probabilistic graph representation of edges. Background Technology

[0002] High-resolution remote sensing images refer to remote sensing images with high spatial resolution acquired by remote sensing carriers such as satellites and drones. They are also an important component of modern remote sensing imagery, characterized by high spatial resolution, high definition, large data volume, multispectral information, and real-time performance. These images can clearly and intuitively present rich details of ground features and the relationships between adjacent features. Currently, semantic segmentation of images is a research hotspot in computer vision. Its essence is category recognition of image regions, i.e., assigning a category label to each pixel in the image. High-resolution remote sensing image semantic segmentation, as an important part of semantic segmentation, can automatically extract surface features from remote sensing images and assign semantic categories to ground targets. High-resolution remote sensing image semantic segmentation has wide applications in disaster assessment and prediction, environmental protection, urban planning, traffic navigation, and military security.

[0003] In recent years, deep learning, especially deep convolutional neural network technology, has experienced rapid development and application. It has demonstrated remarkable feature extraction capabilities in tasks such as image classification, object detection, and semantic segmentation, adaptively extracting both shallow and deep features from images, and exhibiting excellent understanding of complex scenes. Therefore, applying deep learning technology to semantic segmentation of high-resolution remote sensing images has significant practical implications and will bring new development opportunities to remote sensing image processing. However, high-resolution remote sensing images typically consist of large and complex scenes and heterogeneous objects. Furthermore, occlusion and shadow problems caused by lighting conditions and imaging angles during image acquisition lead to poor segmentation performance of existing deep remote sensing segmentation models at object edges. In addition, for small objects, because edge pixels account for a higher proportion of the total pixel count, poor segmentation at the edges also results in poor overall object segmentation. Summary of the Invention

[0004] This invention provides a high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation to overcome the above-mentioned technical problems.

[0005] Semantic segmentation methods for high-resolution remote sensing images based on probabilistic graph-based edge representation include,

[0006] Step 1: Obtain a remote sensing image set and augment the image set. Augmentation involves rotating the remote sensing images in the set at any angle and storing them in the new image set. Then, normalize the images in the augmented image set. Finally, obtain a training set from the normalized image set.

[0007] Step 2: Construct a high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement. This model is used to obtain the semantic segmentation results of the remote sensing image.

[0008] The high-resolution remote sensing image semantic segmentation model includes a convolutional attention unit, a convolutional attention-based encoding / decoding feature extraction module, a probability distribution map unit, and a probability distribution map-based edge enhancement iterative module.

[0009] The convolutional attention unit is used to perform downsampling operations on the input remote sensing image.

[0010] The convolutional attention-based encoding / decoding feature extraction module is used to extract features from the downsampled remote sensing image and output the corresponding feature map.

[0011] The probability distribution map unit is used to convert the feature mapping of the encoding / decoding feature extraction module into a probability map.

[0012] The edge enhancement iterative module based on the probability distribution map is used to iteratively enhance the probability map obtained from the probability distribution map units, and obtain the semantic segmentation result of the remote sensing image based on the iteratively enhanced probability map.

[0013] Step 3: Train the high-resolution remote sensing image semantic segmentation model based on the training set, obtain the remote sensing image to be segmented, input the remote sensing image to be segmented into the trained high-resolution remote sensing image semantic segmentation model, and obtain the semantic segmentation result.

[0014] Preferably, the step of converting the feature map of the encoding / decoding feature extraction module into a probability map includes first applying a softmax activation function to the feature map to obtain a preliminary semantic segmentation result, then using the CV2.Findcontours function to determine the boundary of each segmentation target in the segmentation result and outputting the coordinates of the pixels on each boundary, then using the CV2.PointpolygonTest function to determine the pixels inside and on each boundary and calculate the shortest distance from each pixel to this boundary, and finally using the Sigmoid Alpha Function (SAF) to map the shortest distance from each pixel inside and on each boundary to this boundary to the probability value of this pixel belonging to this segmentation target, which is the target enclosed by this boundary.

[0015] Preferably, the convolutional attention unit is used to perform downsampling operations on the input remote sensing image, which includes obtaining remote sensing images with sizes of 2 / h, 4 / h, 8 / h, and 16 / h from the remote sensing image of size h through a downsampling process.

[0016] Preferably, the SAF function is represented by formula (1).

[0017]

[0018]

[0019] O = max(H) (i,j) );i,j∈β (1)

[0020] Where α is the transformation factor for generating different probability distribution maps; C is a constant used to keep the value range of Sigmoid AlphaFunction within [0,1] and ensure that the probability of the pixel farthest from the boundary is 1; (i,j) represents the coordinate point, H (i,j) β represents the shortest distance from pixel (i,j) to the boundary; O represents the size of the segmentation target; β represents the object being segmented.

[0021] Preferably, the probability graph calculates the loss according to formula (2).

[0022]

[0023]

[0024] Among them, G i Let P be the i-th true value of the probability graph. i Let p be the i-th predicted probability map, and θ be the pixel in the image domain; loss is the loss for the i-th probability map, n is the number of probability maps, and λ is the loss for the i-th probability map. i It is the loss weight of each segmentation target in the probabilistic graph.

[0025] This invention provides a high-resolution semantic segmentation method for remote sensing images based on probabilistic graph-based edge representation. It extracts features from remote sensing images through a convolutional self-attention-based encoding / decoding feature extraction module. The convolutional self-attention unit has a structure similar to a self-attention mechanism, utilizing large-kernel convolution to achieve global dependency similar to self-attention, thus enhancing important information, especially edge regions. Subsequently, an edge enhancement iterative module based on a probability distribution graph further strengthens the segmentation effect on edge regions, improving feature extraction and processing of remote sensing object edge information. This enhances the recognition accuracy of small objects and complex boundary information in remote sensing images, achieving accurate semantic segmentation of remote sensing object edges. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart of the method of the present invention;

[0028] Figure 2 This is a schematic diagram of the remote sensing image semantic segmentation model based on probabilistic graph representation of edges in this invention;

[0029] Figure 3 This is a schematic diagram of the multi-scale fusion encoding and decoding feature extraction module based on convolutional self-attention in this invention.

[0030] Figure 4 This is a schematic diagram of the probability distribution graph unit structure of the present invention;

[0031] Figure 5 This is a schematic diagram of the edge reinforcement iterative module structure based on probability distribution graphs in this invention;

[0032] Figure 6 This is a graph showing the prediction results of the present invention and comparative experiments on the Vaihingen dataset. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0034] Figure 1 This is a flowchart of the method of the present invention, as shown below. Figure 1 As shown, the method in this embodiment may include:

[0035] Step 1: Obtain a remote sensing image set and augment the image set. Augmentation involves rotating the remote sensing images in the set at any angle and storing them in the new image set. Then, normalize the images in the augmented image set. Finally, obtain a training set from the normalized image set.

[0036] Step 2: Construct a high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement. This model is used to obtain the semantic segmentation results of the remote sensing image.

[0037] The high-resolution remote sensing image semantic segmentation model includes a convolutional attention unit, a convolutional attention-based encoding / decoding feature extraction module, a probability distribution map unit, and a probability distribution map-based edge enhancement iterative module.

[0038] The convolutional attention unit is used to perform downsampling operations on the input remote sensing image.

[0039] The convolutional attention-based encoding / decoding feature extraction module is used to extract features from the downsampled remote sensing image and output the corresponding feature map.

[0040] The probability distribution map unit is used to convert the feature mapping of the encoding / decoding feature extraction module into a probability map.

[0041] The edge enhancement iterative module based on the probability distribution map is used to iteratively enhance the probability map obtained from the probability distribution map units, and obtain the semantic segmentation result of the remote sensing image based on the iteratively enhanced probability map.

[0042] Step 3: Train the high-resolution remote sensing image semantic segmentation model based on the training set, obtain the remote sensing image to be segmented, input the remote sensing image to be segmented into the trained high-resolution remote sensing image semantic segmentation model, and obtain the semantic segmentation result.

[0043] Based on the above scheme, features of remote sensing images are extracted through an encoding / decoding feature extraction module based on convolutional self-attention. The convolutional self-attention unit has a structure similar to the self-attention mechanism, utilizing large-kernel convolution to achieve global dependency similar to the self-attention mechanism, thus strengthening important information, especially edge parts. Subsequently, an edge enhancement iterative module based on probability distribution maps enhances the segmentation effect on edge parts, improving feature extraction and processing of remote sensing ground object edge information, enhancing the recognition accuracy of small objects and complex boundary information in remote sensing images, and achieving accurate semantic segmentation of remote sensing ground object edges.

[0044] Specifically, this embodiment provides detailed information on the semantic segmentation method, including:

[0045] Step 1: Obtain the remote sensing image set and augment it. Augmentation involves rotating the remote sensing images in the set at any angle before storing them. To prevent overfitting and improve model robustness, data augmentation typically includes vertical and horizontal flipping, 90° rotation, etc., of the remote sensing images. Normalization is then performed on the augmented images. Normalization is a data preprocessing operation that maps the original remote sensing image data to the range of 0 to 1. The training set is then obtained from the normalized remote sensing image set.

[0046] Step 2: Construct a high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement, as shown in the figure. Figure 2 As shown, the high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement is used to obtain the semantic segmentation results of remote sensing images.

[0047] The high-resolution remote sensing image semantic segmentation model includes a convolutional attention unit, a convolutional attention-based encoding / decoding feature extraction module, a probability distribution map unit, and a probability distribution map-based edge enhancement iterative module.

[0048] The convolutional attention unit is used to downsample the input remote sensing image, replacing the convolutional blocks in the UNet model. This convolutional form can use large kernel convolutions to achieve a global representation capability similar to a self-attention mechanism.

[0049] The convolutional attention-based encoding / decoding feature extraction module is used to extract features from the downsampled remote sensing image and output the corresponding feature map. The structure of the encoding / decoding feature extraction module is as follows: Figure 3 As shown, the encoder-decoder uses the UNet model's encoder-decoder structure and also employs skip connections. The difference is that the convolutional blocks in the encoder are replaced with convolutional self-attention modules. The encoder-decoder feature extraction module extracts features from the input remote sensing image and outputs the corresponding feature maps. The encoder is responsible for extracting high-level semantic features from the input remote sensing image. The encoder network has four layers; as the network deepens, the size of the feature map gradually decreases, while the number of channels in the feature map gradually increases to capture semantic features at different scales. The decoder is responsible for restoring the abstract features extracted by the encoder to the original input image size. The decoder of this invention also consists of four layers, composed of multiple upsampling operations (transposed convolutions). Skip connections are used in the decoder to obtain multi-scale fused feature maps.

[0050] The convolutional attention unit is used to perform downsampling operations on the input remote sensing image, including downsampling a remote sensing image of size h to obtain remote sensing images of sizes 2 / h, 4 / h, 8 / h, and 16 / h respectively. The convolutional self-attention unit can use large kernel convolution to obtain a global representation capability similar to a self-attention mechanism.

[0051] The probability distribution map unit is used to convert the feature mapping of the encoding / decoding feature extraction module into a probability map. The structure of the probability distribution map unit is as follows: Figure 4 As shown, the conversion of the feature map of the encoding / decoding feature extraction module into a probability map includes first applying the softmax activation function to the feature map to obtain a preliminary semantic segmentation result, then using the CV2.Findcontours function to determine the boundary of each segmentation target in the segmentation result and outputting the coordinates of the pixels on each boundary, then using the CV2.PointpolygonTest function to determine the pixels inside and on each boundary and calculate the shortest distance from each pixel to this boundary, and finally using the Sigmoid Alpha Function (SAF) to map the shortest distance from each pixel inside and on each boundary to this boundary to the probability value of this pixel belonging to this segmentation target. This segmentation target is the target enclosed by this boundary. Specifically, the SAF function is a variant of the Sigmoid function. The Sigmoid function can map distance values ​​to the range [0,1], which can be a probability value. The SAF function is represented by formula (1).

[0052]

[0053]

[0054] O = max(H) (i,j) );i,j∈β (1)

[0055] Where α is the transformation factor for generating different probability distribution maps; C is a constant used to keep the value range of Sigmoid AlphaFunction within [0,1] and ensure that the probability of the pixel farthest from the boundary is 1; (i,j) represents the coordinate point, H (i,j) β represents the shortest distance from pixel (i,j) to the boundary; O represents the size of the segmentation target; β represents the object being segmented.

[0056] The probability graph calculates the loss according to formula (2).

[0057]

[0058]

[0059] Among them, Gi Let P be the i-th true value of the probability graph. i Let p be the i-th predicted probability map, and θ be the pixel in the image domain; loss is the loss for the i-th probability map, n is the number of probability maps, and λ is the loss for the i-th probability map. i It is the loss weight of each segmentation target in the probabilistic graph.

[0060] The edge enhancement iterative module based on the probability distribution map is used to iteratively enhance the probability map obtained from the probability distribution map unit. The structure of the edge enhancement iterative module is as follows: Figure 5 As shown, the semantic segmentation result of the remote sensing image is obtained based on the probability map after iterative enhancement. This iterative module can make full use of the prediction information of the low-level iteration to improve the accuracy of the prediction of the high-level iteration, and can optimize the parameters of the high-level iteration layer by backpropagating the gradient of the high-level iteration layer. The mathematical expression of the iterative module can be expressed by formula (3):

[0061] f1=ConvBlock(f0); P1=σ(f1)

[0062]

[0063] P i =σ(f i );i∈(i,n] (3)

[0064] in This refers to a concatenated operation, where σ(·) is the SAF function, and f0 represents the feature fusion map obtained from the encoding / decoding feature extraction module. f0 is input into the first layer of the iterative module to generate feature map f1. Then, f1 is processed by the SAF function to generate a probability map, which is then fed into the next iterative layer to provide prior information for generating the second probability map. In this way, a set of probability map predictions can be obtained.

[0065] Specifically, the iterative module uses the SAF function to model the relationship between the probability of a pixel belonging to a target pixel and its distance to the target boundary based on the shortest distance from each pixel in the segmented target. This relationship is then continuously refined through the iterative module, generating a series of probability distribution maps. Due to boundary blurring and partial boundary occlusion, object boundaries are often difficult to segment. A single probability map cannot accurately describe the probability distribution of target pixels within the boundary. To accommodate multiple different probability distribution maps, an iterative module is proposed. This module can fully utilize the prediction information from lower-level iterations to improve the accuracy of higher-level iterations, and can optimize the parameters of higher-order iteration layers by backpropagating gradients. Furthermore, to better segment edge regions, asymmetric convolutional blocks As Conv (1×3 and 3×1) are used in the iterative module to better highlight the edges of different objects.

[0066] Step 3: Train the high-resolution remote sensing image semantic segmentation model based on the training set, obtain the remote sensing image to be segmented, input the remote sensing image to be segmented into the trained high-resolution remote sensing image semantic segmentation model, and obtain the semantic segmentation result.

[0067] This embodiment uses real remote sensing image data for experiments. Two sets of publicly available real remote sensing image datasets are used to test and illustrate the high-resolution remote sensing image semantic segmentation model based on probabilistic graph edge representation provided by this invention, as well as to analyze and evaluate its application effects.

[0068] 1. Dataset and parameter settings

[0069] This embodiment uses two publicly available high-resolution remote sensing image datasets (Potsdam dataset and Vaihingen dataset) from the ISPRS2D semantic annotation competition for experiments and analysis. The datasets employ digital surface models (DSMs) generated from high-resolution orthogonal photographs and corresponding dense image matching techniques.

[0070] The ISPRS Vaihingen dataset contains 33 images of varying sizes, with an average size of 2494×2064 pixels and a spatial resolution of 9 cm. The images include three bands: near-infrared (NIR), red (R), and green (G). The labels include six categories: impermeable surfaces, buildings, low-lying plants, trees, cars, and others. Twelve images (numbered 1, 4, 7, 10, 13, 16, 19, 22, 25, 28, 31, and 33) were used for testing, image number 30 was used for validation, and the remaining 20 images were used for training. The images were cropped into 256×256 pixel patches.

[0071] The ISPRS Potsdam dataset contains 38 images, each 6000×6000 pixels in size, with a spatial resolution of 5 cm, using three bands: red (R), green (G), and blue (B). The label categories and number are consistent with the Vaihingen dataset. This embodiment uses images numbered 2_13, 2_14, 3_13, 3_14, 4_13, 4_14, 4_15, 5_13, 5_14, 5_15, 6_13, 6_14, 6_15, and 7_13 for testing, image number 2_10 for validation, and the remaining 22 images (excluding image 7_10 with incorrect annotations) for training. The experiment only utilizes the red, green, and blue bands, cropping the original image patches into 256×256 pixel blocks.

[0072] The batch size for training the model in this invention is set to 16, with each training session consisting of 300 epochs. The learning rate is dynamically adjusted using cosine annealing, with an initial learning rate of 1e-3, a learning rate decay coefficient of 0.2, and a learning rate decay interval of 5. The AdamW optimizer is used to optimize the parameters.

[0073] 2. Experimental Evaluation Indicators

[0074] Overall accuracy (OA) is a performance metric used to evaluate classification models. In image semantic segmentation tasks, it refers to the proportion of correctly classified pixels out of the total number of pixels. Its calculation formula is:

[0075]

[0076] F1 score and mF1 score are metrics for measuring the performance of classification models and are often used to evaluate the accuracy of binary or multi-class classification models.

[0077] Intersection over Union (IoU) and Mean Intersection over Union (mIoU) are commonly used metrics for measuring the performance of object detection and semantic segmentation models. The calculation formulas are:

[0078]

[0079] 3. Analysis and Evaluation of Experimental Results

[0080] The results of experiments using two sets of remote sensing image data for a remote sensing image semantic segmentation model based on probabilistic graph edge representation provided in this embodiment are shown in Tables 1 and 2. Figure 6 This is a graph showing the prediction results of our method and comparative experiments on the Vaihingen dataset.

[0081] Table 1. Comparison Experiment Results of the Vaihingen Dataset

[0082]

[0083]

[0084] Table 2. Comparison of Potsdam dataset experimental results

[0085]

[0086] The experiment introduced FCN and PSPNet, models based on DCNN, and Segmenter and UNetFormer, improved models based on Transformer. Based on the classification results, the following conclusions can be drawn:

[0087] As can be seen from the table, the present invention achieves better results in semantic segmentation models. Comparative tests show that the proposed model achieves better performance in terms of mIoU, mF1, and OA values ​​compared to improved DCNN-based models based on FCN and PSPNet, and improved Transformer-based models based on Segmenter and UNetFormer.

[0088] Overall beneficial effects:

[0089] This invention provides a high-resolution semantic segmentation method for remote sensing images based on probabilistic graph-based edge representation. It extracts features from remote sensing images through a convolutional self-attention-based encoding / decoding feature extraction module. The convolutional self-attention unit has a structure similar to a self-attention mechanism, utilizing large-kernel convolution to achieve global dependency similar to self-attention, thus enhancing important information, especially edge regions. Subsequently, an edge enhancement iterative module based on a probability distribution graph further strengthens the segmentation effect on edge regions, improving feature extraction and processing of remote sensing object edge information. This enhances the recognition accuracy of small objects and complex boundary information in remote sensing images, achieving accurate semantic segmentation of remote sensing object edges.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation, characterized in that, include, Step 1: Obtain a remote sensing image set and augment the image set. Augmentation involves rotating the remote sensing images in the set at any angle and storing them in the new image set. Then, normalize the images in the augmented image set. Finally, obtain a training set from the normalized image set. Step 2: Construct a high-resolution remote sensing image semantic segmentation model based on probabilistic graph distribution edge enhancement. This model is used to obtain the semantic segmentation results of the remote sensing image. The high-resolution remote sensing image semantic segmentation model includes a convolutional attention unit, a convolutional attention-based encoding / decoding feature extraction module, a probability distribution map unit, and a probability distribution map-based edge enhancement iterative module. The convolutional attention unit is used to perform downsampling operations on the input remote sensing image. The convolutional attention-based encoding / decoding feature extraction module is used to extract features from the downsampled remote sensing image and output the corresponding feature map. The probability distribution map unit is used to convert the feature mapping of the encoding / decoding feature extraction module into a probability map. The edge enhancement iterative module based on the probability distribution map is used to iteratively enhance the probability map obtained from the probability distribution map units, and obtain the semantic segmentation result of the remote sensing image based on the iteratively enhanced probability map. Step 3: Train the high-resolution remote sensing image semantic segmentation model based on the training set, obtain the remote sensing image to be segmented, input the remote sensing image to be segmented into the trained high-resolution remote sensing image semantic segmentation model, and obtain the semantic segmentation result.

2. The high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation according to claim 1, characterized in that, The process of converting the feature map of the encoding / decoding feature extraction module into a probabilistic map includes: first, applying the softmax activation function to the feature map to obtain preliminary semantic segmentation results; then, using the CV2.Findcontours function to determine the boundary of each segmentation target in the segmentation results and outputting the coordinates of the pixels on each boundary; next, using the CV2.PointpolygonTest function to determine the pixels inside and on each boundary and calculate the shortest distance from each pixel to this boundary; finally, using the Sigmoid Alpha Function (SAF) to map the shortest distance from each pixel inside and on each boundary to this boundary to the probability value of this pixel belonging to this segmentation target, which is the target enclosed by this boundary.

3. The high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation according to claim 1, characterized in that, The convolutional attention unit is used to perform downsampling operations on the input remote sensing image, including obtaining remote sensing images with sizes h / 2, h / 4, h / 8, and h / 16 from a remote sensing image of size h through a downsampling process.

4. The high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation according to claim 2, characterized in that, The SAF function is represented by formula (1). (1) Where α is the transformation factor for generating different probability distribution maps; C is a constant used to keep the value range of Sigmoid AlphaFunction within [0,1] and ensure that the probability of the pixel farthest from the boundary is 1. Represents the coordinates of the point. For pixels The shortest distance to the boundary; O Indicates the size of the segmentation target; β It represents the object being segmented.

5. The high-resolution remote sensing image semantic segmentation method based on probabilistic graph edge representation according to claim 1, characterized in that, The probability graph calculates the loss according to formula (2). (2) in, For the probability graph of the th One true value For the first A predicted probability map, p Represents the pixel in the image domain ; loss For the first The loss of a probability graph, n The number of probability graphs. It is the loss weight of each segmentation target in the probabilistic graph.

Images