A point cloud semantic segmentation method for intelligent driving of new energy vehicles

Abstract

This invention relates to the field of intelligent driving technology, specifically to a point cloud semantic segmentation method for intelligent driving of new energy vehicles. It constructs a lightweight PointPsWE network model, utilizes an embedding module to perform high-dimensional mapping on the original point cloud data, and then performs feature extraction and aggregation through multiple cascaded processing stages. Each processing stage includes a spatial attention and angle transformation module for extracting local features, and a PSWE module for feature fusion and dimensionality reduction through hybrid pooling and point convolution. Finally, the semantic segmentation result is output through a classification head and feature propagation module. This method effectively reduces the number of model parameters and computational complexity, reduces memory usage and inference time, and maintains competitive classification and segmentation accuracy on datasets such as ModelNet40, ScanObjectNN, and S3DIS. It solves the technical problem that existing methods are difficult to deploy efficiently in automotive embedded devices due to high computational complexity and large number of parameters.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology, and in particular to a point cloud semantic segmentation method for intelligent driving of new energy vehicles. Background Technology

[0002] Autonomous driving systems rely on sensors such as LiDAR to perceive the 3D environment, and point clouds have become the most common data type for representing the 3D information of objects. Compared with image data, point cloud data has three significant characteristics: unstructured, permutation invariant, and sparsity. Unstructured means that a point cloud is a collection of points in 3D space, belonging to irregular data; permutation invariance means that rotation of the point cloud in space does not affect its internal structure and properties; sparsity is manifested in the uneven distribution of points in the point cloud and the existence of missing points. Therefore, the key to point cloud data analysis algorithms lies in adapting to its data characteristics and achieving effective feature extraction and aggregation, and its performance directly affects the perception effect of autonomous driving systems.

[0003] Currently, deep learning-based point cloud processing methods mainly fall into three categories: multi-view methods, voxel-based methods, and direct point-based methods. Multi-view and voxel-based methods convert point clouds into 2D images and 3D meshes, respectively, but these methods are prone to losing original data features during format conversion. Direct point-based methods, on the other hand, process point clouds directly through operations such as multilayer perceptrons, convolution, graph convolution, or attention mechanisms, exhibiting excellent performance and becoming the mainstream research direction. However, these methods generally suffer from high computational complexity, large memory consumption, and a large number of model parameters, making them difficult to deploy efficiently in resource-constrained environments such as automotive embedded devices.

[0004] Therefore, there is still a lack of lightweight methods in the existing technology that can significantly reduce the computational complexity and memory usage of the model while ensuring the accuracy of point cloud semantic segmentation. Summary of the Invention

[0005] The purpose of this invention is to provide a point cloud semantic segmentation method for intelligent driving of new energy vehicles, which solves the problems of existing methods having a large number of model parameters, high computational complexity, and difficulty in maintaining high segmentation accuracy while being applicable to the resource-constrained environment of in-vehicle embedded devices.

[0006] To achieve the above objectives, this invention provides a point cloud semantic segmentation method for intelligent driving of new energy vehicles, comprising the following steps: Obtain the raw point cloud data of the target scene; The raw point cloud data is input into a pre-trained PointPswe network model, which includes an embedding module, multiple cascaded processing stages, a classification head, and a feature propagation module. Each processing stage includes a spatial attention and angle transformation module and a PSWE module. The Embedding module performs high-dimensional feature mapping on the input raw point cloud data; The high-dimensional features after mapping are extracted and aggregated step by step through multiple cascaded processing stages. In each processing stage, the spatial attention and angle transformation module performs local feature extraction on the input features, and the PSWE module performs aggregation and feature fusion on the extracted local features. The aggregated features output by the last PSWE module are processed through the classification head and the feature propagation module to finally output the semantic segmentation result of the point cloud.

[0007] The spatial attention and angle transformation module performs local feature extraction on the input features, specifically including: First, perform a multi-scale adaptive radius grouping operation to determine the neighborhood of each center point; Then, position encoding, spatial attention weighting, and feature angle transformation are sequentially performed on the features within the neighborhood; Finally, the inputs of the spatial attention and angle transformation modules are added to the processed outputs through residual connections.

[0008] First, a multi-scale adaptive radius grouping operation is performed to determine the neighborhood of each center point, specifically including: The ball query method is used to search for neighboring points of each point as the center; Calculate the Euclidean distance between neighboring points and the center point, and adaptively determine the sphere query radius based on the distance sorting results and the preset number of samples; By combining a multi-scale grouping strategy, multiple adaptive radii of different scales are used to group the features, and the features extracted at each scale are connected to form multi-scale features.

[0009] Specifically, the spatial attention weighting includes: The input features are processed using dilated convolutional layers to generate single-channel feature maps. Batch normalization is performed on the single-channel feature map; The normalized feature map is converted into spatial attention weights using the Sigmoid function; The spatial attention weights are element-wise multiplied with the original input features to obtain the spatial attention features.

[0010] Specifically, the feature angle transformation includes: The input containing location and feature information is converted into two angular parameters through two-dimensional convolution, layer normalization, and activation functions. The input features are projected onto different directions defined by the angle parameters to generate feature representations in multiple directions; Connect and aggregate the feature representations of the multiple directions; One-dimensional convolution is used to adjust the number of channels in the aggregated features to match the number of input features.

[0011] The PSWE module performs aggregation and feature fusion on the extracted local features, specifically including: The input features are aggregated by a hybrid pooling layer, which is a combination of PSWE pooling and Max pooling, and its output is a weighted sum of the results of the two. The aggregated features are fused and channel transformed by point convolutional layers, which employ depthwise separable convolution and inverted bottleneck design.

[0012] The PSWE pooling process involves calculating the Sliced-Wasserstein distance between the sample feature distribution and the reference distribution and then aggregating the features accordingly. The Max pooling process is as follows: select the largest feature value from each pooling window as the output.

[0013] Specifically, the aggregated features output by the last PSWE module are processed through the classification head and the feature propagation module to finally output the semantic segmentation result of the point cloud, including: The feature propagation module achieves upsampling through an interpolation function and performs channel dimension transformation and feature fusion through a multilayer perceptron to output the final segmentation result.

[0014] This invention discloses a point cloud semantic segmentation method for intelligent driving of new energy vehicles. It constructs a lightweight PointPsWE network model, utilizes an embedding module to perform high-dimensional mapping on the original point cloud data, and then performs feature extraction and aggregation through multiple cascaded processing stages. Each processing stage includes a spatial attention and angle transformation module for extracting local features, and a PSWE module for feature fusion and dimensionality reduction through hybrid pooling and point convolution. Finally, a classification head and feature propagation module output the semantic segmentation result. The spatial attention and angle transformation module combines multi-scale adaptive radius grouping, spatial attention weighting, and feature angle transformation. The PSWE module employs hybrid pooling combining PSWE pooling and max pooling, as well as point convolutional layers with depthwise separable convolution and an inverted bottleneck design. This effectively reduces the number of model parameters and computational complexity, decreases memory usage and inference time, while maintaining competitive classification and segmentation accuracy on datasets such as ModelNet40, ScanObjectNN, and S3DIS. It solves the technical problem that existing methods are difficult to deploy efficiently in automotive embedded devices due to high computational complexity and large number of parameters. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0016] Figure 1 This is the overall framework diagram of PointPswe of the present invention.

[0017] Figure 2 This is a schematic diagram of the multi-scale and adaptive radius algorithm of the present invention.

[0018] Figure 3 This is a schematic diagram of the SA layer structure of the present invention.

[0019] Figure 4 This is a structural diagram of the AT layer of the present invention.

[0020] Figure 5 This is a diagram showing the 3D object classification results in ScanObjectNN and ModelNet40 of this invention.

[0021] Figure 6 This is a diagram showing the 3D semantic segmentation results on the S3DIS of this invention.

[0022] Figure 7 This is a detailed semantic segmentation diagram of PointPswe in region 5 of S3DIS of the present invention.

[0023] Figure 8This is a visualization of the S3DIS semantic segmentation results of the present invention.

[0024] Figure 9 This is a flowchart of the point cloud semantic segmentation method for intelligent driving of new energy vehicles according to the present invention. Detailed Implementation

[0025] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0026] The first embodiment of this application is as follows: Please see Figures 1 to 9 ,in, Figure 1 This is the overall framework diagram of PointPswe of the present invention. Figure 2 This is a schematic diagram of the multi-scale and adaptive radius algorithm of the present invention. R is the calculated adaptive radius, r is the initial radius of the ball query algorithm, and a is the radius scaling factor of the MSG algorithm. Figure 3 This is a schematic diagram of the SA layer structure of the present invention. Figure 4 This is a structural diagram of the AT layer of this invention. Angle transformation converts the input features based on their positional information, generating feature representations from different angles. Conv1d is a simple one-dimensional convolution used for channel transformation. Figure 5 This is a diagram showing the 3D object classification results in ScanObjectNN and ModelNet40 of this invention. Figure 6 This is a diagram showing the 3D semantic segmentation results on the S3DIS of this invention. Figure 7 This is a detailed semantic segmentation diagram of PointPswe in region 5 of S3DIS of the present invention. Figure 8 This is a visualization of the S3DIS semantic segmentation results of the present invention. Figure 9 This is a flowchart of the point cloud semantic segmentation method for intelligent driving of new energy vehicles according to the present invention.

[0027] This invention provides a point cloud semantic segmentation method for intelligent driving of new energy vehicles, comprising the following steps: S101: Obtain the raw point cloud data of the target scene; S102: Input the raw point cloud data into a pre-trained PointPswe network model. The PointPswe network model includes an Embedding module, multiple cascaded processing stages, a classification head, and a feature propagation module. Each processing stage includes a spatial attention and angle transformation module and a PSWE module. S103: The Embedding module performs high-dimensional feature mapping on the input raw point cloud data; S104: The mapped high-dimensional features are extracted and aggregated step by step through multiple cascaded processing stages. In each processing stage, the spatial attention and angle transformation module performs local feature extraction on the input features, and the PSWE module performs aggregation and feature fusion on the extracted local features. S105: The aggregated features output by the last PSWE module are processed through the classification head and the feature propagation module to finally output the semantic segmentation result of the point cloud.

[0028] Specifically, the network structure of PointPswe is established, as follows: Figure 1 As shown. The model consists of four stages, each consisting of a SAAT module and a PSWE module. Its principle can be explained by the following formula: In this formula, This indicates the output. It is an aggregation function, which is a hybrid pooling operation consisting of PAWE and max pooling. This represents a hybrid pooling operation, which performs finer feature aggregation. Due to the irregularity of point cloud data format, the geometry of different local regions is sparse and irregular. This necessitates designing suitable local extractors to capture various geometric features. SAAT stands for Spatial Attention and Angle Transformation Module. , and Representing points respectively Input features, points Location and point The location. The general architecture of PointPswe is as follows: Figure 1 As shown, PointPswe uses a U-Net architecture, which includes three stages: encoding, decoding, and classification / segmentation processing.

[0029] Embedding module: such as Figure 1 As shown, the Embedding module consists of one-dimensional convolution, batch normalization, and activation functions. It can map the input point cloud to a higher-dimensional space, allowing subsequent models to process and understand the data more effectively.

[0030] Multi-scale adaptive radius algorithm: such as Figure 1 As shown, the multi-scale adaptive algorithm is located in Grouping within the SAAT Block. Due to the unordered nature of point cloud data, local aggregation must explicitly model the spatial relationships between points.

[0031] Common strategies include K-nearest neighbors, sphere query, and grid partitioning. Sphere query is chosen as the grouping method. Sphere query selects all points within a given radius as neighbors, allowing for flexible definition of the neighborhood size to accommodate point clouds of varying densities. A key aspect of the sphere query method is setting an appropriate radius value. There is a close relationship between the radius value and the receptive field of the neural network. Therefore, a multi-scale adaptive radius calculation method is designed. The calculation process of the adaptive radius method is shown in the formula.

[0032] The first step is to calculate the Euclidean distance between adjacent points and the center point. Wherein... This represents the Euclidean distance between neighboring points and the center point; , and Representing the neighboring points respectively The three-dimensional coordinates, that is, the spatial location information of neighboring points collected by the lidar; , and They represent the center points respectively. The three-dimensional coordinates, that is, the spatial location information of the center of the local area determined after sampling.

[0033] The second step is to construct a distance matrix. This matrix stores the distance information between the query point and the center point. The argsort algorithm is then used to sort the distances in ascending order within the distance matrix, and the sorted indices are returned. This represents the index sequence after distance sorting.

[0034] Finally, the adaptive radius related to the query point and the center point is calculated. The adaptive radius value is k, which is the number of samples in the ball query algorithm group and is a hyperparameter.

[0035] Multi-scale is a multi-scale grouping (MSG) proposed by PointNet++. It applies grouped layers with different radius values ​​and then aggregates features extracted at each scale. Features from different scales are connected to form multi-scale features. This combines multi-scale grouping with adaptive radii. In the MSG algorithm, the initial radius and radius scaling factor are hyperparameters. Figure 2As shown, MSG controls the ball query algorithm to extract features at multiple scales by setting the initial radius and radius scaling factor. The adaptive radius algorithm calculates and sets an appropriate radius value based on the ball query algorithm to extract features. By combining the adaptive radius with multi-scale features, local and global features of point cloud data can be effectively captured, enhancing the model's expressiveness. Through design, this model can not only adapt to different local structures but also extract useful information from multiple scales.

[0036] Spatial feature angle transformation module: such as Figure 1 As shown, the spatial feature angle transformation module is the SAAT Block. It consists of a Position layer, an SA layer, an AT layer, a Result layer, and residual connections.

[0037] The Position layer plays a crucial role in the attention mechanism. Using 2D convolution and normalization operations, it transforms raw positional coordinate information into channel-dimensional feature representations. It aims to integrate positional information, facilitate feature dimensional matching, and enhance the model's understanding of spatial relationships.

[0038] SA layer. The location of the SA layer in the model is as follows: Figure 1 As shown. It can introduce a spatial attention mechanism into the model to enhance feature extraction capabilities. Its detailed structure is as follows. Figure 3 As shown. The SA layer processes the input data X to obtain the feature Y containing spatial information. This process can be divided into the following two steps: (1) Generating Spatial Attention Weights. The input data X is processed through a 7×7 convolutional layer to produce single-channel features. The generated features are batch normalized to adjust the feature distribution for improved stability. Then, the feature map is converted into attention weights using the sigmoid function, outputting values ​​between 0 and 1. In particular, the 7×7 convolution used here is a dilated convolution. Dilated convolutions increase the receptive field by inserting gaps between the elements of the convolutional kernel, enabling it to capture a wider range of contextual information without increasing the number of parameters.

[0039] (2) Generate spatial attention features. For example... Figure 3 As shown, the previously obtained spatial attention weights are multiplied by the input data X, and Y represents the spatial attention features. The formula is as follows: where y represents the spatial attention features, and x represents the input data. This represents the spatial attention weights.

[0040] The calculated attention weights are applied to the original feature X by performing element-wise multiplication. This method suppresses irrelevant information and enhances features in specific regions (high-weighted areas), providing more contextual information for subsequent processing. By focusing on important regions, spatial attention mechanisms can help reduce the complexity of subsequent computations.

[0041] AT layer. The location of the AT layer in the model is as follows: Figure 1 As shown. It can perform vector angle transformations on features. Detailed structure is as follows. Figure 4 As shown. The AT layer mainly consists of two parts: angle transformation and feature information aggregation. Its implementation steps are as follows: (1) X consists of original location information and feature information. X is converted into two feature maps using two-dimensional convolution, layer normalization and activation function. The output of the feature map is normalized to between 0 and 1 using sigmoid activation function and multiplied by π to generate angle1 and angle2.

[0042] (2) Apply simple 2D convolution to the input features to generate new features. Using trigonometric functions Combined with angular information, a new feature representation is formed. The formula is as follows: Where... Represented as a new feature, For trigonometric functions, For angle The arcsine function value, For angle The value of the inverse cosine function, For angle The arcsine function value, For angle The value of the inverse cosine function, The polar angle of the eigenvector. The azimuth angle of the eigenvector.

[0043] This formula will Projected to and In the defined direction, features that transform in different directions are connected to form a new feature representation in three-dimensional space.

[0044] (3) Use the reshape function to transform features. The data format is as follows. A flag parameter is set. Its function is: when flag=1, sum along the last dimension; otherwise, take the maximum value of the last dimension. Finally, a one-dimensional convolution is used to change the number of channels from three to the original number of channels.

[0045] Finally, residual connections are used to directly add the model's original input X to the output. Residual connections can effectively alleviate the gradient vanishing problem.

[0046] PSWE and Max hybrid pooling: such as Figure 1 As shown, the PSWE Block consists of an aggregation layer, a hybrid pooling layer of PSWE and Max, and a point convolutional layer.

[0047] PSWE pooling calculates the Sliced-Wasserstein distance (SWD). The SWD is used to compare the sample feature distribution with a reference distribution. The SWD is calculated based on the ranking results. Finally, features are aggregated, combining the features of each sample into the final representation. Max pooling selects the maximum eigenvalue from each pooling window as the output, reducing the dimensionality of the feature map while preserving global feature information.

[0048] In PSWE Block, a hybrid pool is used, such as... Figure 1 As shown. Hybrid pooling is a combination of PSWE pooling and max pooling, and it can be expressed by the following formula: in, It is a random number between 0 and 1. It indicates the random choice of max pooling and average pooling for each pooling operation. The parallel combination of max pooling and PSWE pooling can simultaneously preserve global features (from max pooling) and locally important information (from PSWE pooling). Hybrid pooling can enhance robustness to some extent and prevent network overfitting. Furthermore, this method requires relatively little computational overhead.

[0049] Considering the issues of model parameter count and feature extraction capability, pointwise convolutional layers are introduced, such as... Figure 1 As shown in Pwconv in PSWEBlock, a depthwise separable convolutional layer is added between the input and output layers. An inverted bottleneck design is employed, increasing the output channels of the intermediate layer by a factor of four to change the number of channels and perform feature fusion. Pointwise convolutional layers can significantly reduce computational complexity while maintaining feature extraction capabilities.

[0050] Classification and segmentation task implementation: such as Figure 1 As shown, the classification head classifies the point cloud features and outputs the classification results. The Feat Propagation module consists of Mlps and Interpolate blocks. The Interpolate block uses the interpolation function to perform upsampling. The MLP performs channel dimension transformation, and the Feat Propagation module outputs the segmentation results.

[0051] After processing point cloud data using the aforementioned network structure, to verify its classification and segmentation performance, the performance of the PointPswe model on the ModelNet40, ScanObjectNN, and S3DIS datasets was evaluated and analyzed. Due to experimental bias, each network model underwent three consecutive training and testing cycles. The best value was used as the final experimental result.

[0052] The 3D object classification results in ScanObjectNN and ModelNet40 are shown below. Figure 5 : The PointPswe model's classification performance on the ModelNet40 and ScanObjectNN datasets is reported. The model parameter numbers (#Params) and overall accuracy (%) are listed. Bold values ​​indicate best results, and underlined values ​​indicate second-best results.

[0053] The 3D semantic segmentation results on S3DIS are shown below. Figure 6 : Semantic segmentation on S3DIS Area 5. Bold values ​​indicate the best results, and underlined values ​​indicate the second-best results. All comparison algorithms are single-modal point cloud data processing algorithms, without a voting strategy.

[0054] Figure 7 Detailed semantic segmentation of PointPswe in region 5 of S3DIS. Bold values ​​indicate the best results, and underlined values ​​indicate the second-best results.

[0055] Figure 8 Visualization results of S3DIS semantic segmentation. Colored blocks correspond to the segmented objects. Ground Truth represents the standard reference segmentation result.

[0056] This invention effectively combines spatial attention mechanisms and feature angle transformation by employing key lightweight structures such as the aforementioned spatial attention and angle transformation modules, PSWE module and its hybrid pooling, depthwise separable convolution, and inverted bottleneck design. This achieves more refined feature extraction and aggregation, significantly reducing the number of model parameters and computational cost, thereby reducing memory usage and inference time. Furthermore, it achieves competitive classification and segmentation accuracy on datasets such as ModelNet40, ScanObjectNN, and S3DIS. In summary, this invention achieves significant progress in lightweight design and efficiency improvement while maintaining competitive accuracy, making it suitable for resource-constrained environments such as automotive embedded devices. It provides an efficient and reliable point cloud semantic segmentation solution for intelligent driving systems in new energy vehicles.

[0057] The above-disclosed embodiments are merely one or more preferred embodiments of this application and should not be construed as limiting the scope of this application. Those skilled in the art can understand that implementing all or part of the above embodiments and making equivalent changes in accordance with the claims of this application still fall within the scope of this application.

Claims

1. A point cloud semantic segmentation method for intelligent driving of new energy vehicles, characterized in that, Includes the following steps: Obtain the raw point cloud data of the target scene; The raw point cloud data is input into a pre-trained PointPswe network model, which includes an embedding module, multiple cascaded processing stages, a classification head, and a feature propagation module. Each processing stage includes a spatial attention and angle transformation module and a PSWE module. The Embedding module performs high-dimensional feature mapping on the input raw point cloud data; The high-dimensional features after mapping are extracted and aggregated step by step through multiple cascaded processing stages. In each processing stage, the spatial attention and angle transformation module performs local feature extraction on the input features, and the PSWE module performs aggregation and feature fusion on the extracted local features. The aggregated features output by the last PSWE module are processed through the classification head and the feature propagation module to finally output the semantic segmentation result of the point cloud. 2.The point cloud semantic segmentation method for intelligent driving of new energy vehicles according to claim 1, characterized in that, The spatial attention and angle transformation module performs local feature extraction on the input features, specifically including: First, perform a multi-scale adaptive radius grouping operation to determine the neighborhood of each center point; Then, position encoding, spatial attention weighting, and feature angle transformation are sequentially performed on the features within the neighborhood; Finally, the inputs of the spatial attention and angle transformation modules are added to the processed outputs through residual connections. 3.The point cloud semantic segmentation method for intelligent driving of new energy vehicles according to claim 2, characterized in that, First, perform a multi-scale adaptive radius grouping operation to determine the neighborhood of each center point, specifically including: The ball query method is used to search for neighboring points of each point as the center; Calculate the Euclidean distance between neighboring points and the center point, and adaptively determine the sphere query radius based on the distance sorting results and the preset number of samples; By combining a multi-scale grouping strategy, multiple adaptive radii of different scales are used to group the features, and the features extracted at each scale are connected to form multi-scale features. 4.The point cloud semantic segmentation method for intelligent driving of new energy vehicles according to claim 2, characterized in that, The spatial attention weighting specifically includes: The input features are processed using dilated convolutional layers to generate single-channel feature maps. Batch normalization is performed on the single-channel feature map; The normalized feature map is converted into spatial attention weights using the Sigmoid function; The spatial attention weights are element-wise multiplied with the original input features to obtain the spatial attention features. 5.The point cloud semantic segmentation method for intelligent driving of new energy vehicles according to claim 2, characterized in that, The feature angle transformation specifically includes: The input containing location and feature information is converted into two angular parameters through two-dimensional convolution, layer normalization, and activation functions. The input features are projected onto different directions defined by the angle parameters to generate feature representations in multiple directions; Connect and aggregate the feature representations of the multiple directions; One-dimensional convolution is used to adjust the number of channels in the aggregated features to match the number of input features. 6.The point cloud semantic segmentation method for intelligent driving of new energy vehicles according to claim 1, characterized in that, The PSWE module performs aggregation and feature fusion on the extracted local features, specifically including: The input features are aggregated by a hybrid pooling layer, which is a combination of PSWE pooling and Max pooling, and its output is a weighted sum of the results of the two. The aggregated features are fused and channel transformed by point convolutional layers, which employ depthwise separable convolution and an inverted bottleneck design.

7. The point cloud semantic segmentation method for intelligent driving of new energy vehicles as described in claim 6, characterized in that, The PSWE pooling process is as follows: calculate the Sliced-Wasserstein distance between the sample feature distribution and the reference distribution, and aggregate features accordingly; The Max pooling process is as follows: select the largest feature value from each pooling window as the output.

8. The point cloud semantic segmentation method for intelligent driving of new energy vehicles as described in claim 1, characterized in that, The aggregated features output by the last PSWE module are processed through the classification head and the feature propagation module to finally output the semantic segmentation result of the point cloud, specifically including: The feature propagation module achieves upsampling through an interpolation function and performs channel dimension transformation and feature fusion through a multilayer perceptron to output the final segmentation result.

Images