| Strengthening the analysis and application of big scene data is beneficial for advancing the development of semantic segmentation technology and providing more AI services for people.However,previous work mostly applied deep learning algorithms to small and medium-scale data,and research on semantic understanding of big scene data is ongoing.Semantic segmentation algorithms for big scene data can segment the edges of different objects in the environment,thereby helping computers understand the environmental features of the real world like humans.Currently,such algorithms have been widely applied in navigation systems,remote sensing mapping,and autonomous driving.This article provides a brief summary and analysis of the development of semantic segmentation for big scene data in recent years,and selects two parts for in-depth exploration: semantic segmentation of large-scale 3D point cloud data and cloud detection for remote sensing images.This thesis proposes a semantic segmentation algorithm model for large-scale point cloud data based on Point Transformer.The model adopts a U-Net-like encoding and decoding structure,composed of MLP modules,Transition-down modules,Transformer modules,and Transition-up modules,with skip connections between the encoder and decoder to improve network segmentation performance.In the Transition-down module,Random Sampling is used to reduce the time cost and memory consumption of the input point cloud down sampling process.To address the problem of key point information loss,Attentive Pooling is proposed to automatically learn and retain useful information.On the Sensat Urban point cloud dataset,the mean intersection over union(m Io U)of the network in this article is 5.7% higher than that of the Point Transformer network,and on the Semantic3 D point cloud dataset,the m Io U is improved by 12.3%.At the same time,both segmentation speed and accuracy are improved.The experimental results show that the proposed network can quickly and accurately perform semantic segmentation of large-scale point cloud data.For the cloud detection in remote sensing images,this article proposes a cloud detection algorithm structure based on a dual attention mechanism.The network also adopts an encoder-decoder structure with skip connections,and adds a dual attention mechanism module between the encoder’s output and the decoder’s input.The dual attention mechanism consists of two parts: a spatial attention mechanism module and a channel attention mechanism module.The network is tested on the 38-Cloud dataset,with a segmentation accuracy of 97.54%,which is 2.22% higher than that of the U-Net network.The experimental results show that the proposed network can accurately segment the rough outline of cloud pixels in remote sensing images. |
