Research On Lightweight Remote Sensing Image Semantic Segmentation Model Based On DeeplabV3+

Remote sensing image semantic segmentation is a key technology for extracting useful information from remote sensing images and has important applications in various fields such as urban planning,agricultural monitoring,disaster warning,and defense.With the rapid development of computer science technology,many traditional semantic segmentation methods have been replaced by deep learning semantic segmentation methods.Deep learning-based remote sensing image semantic segmentation is currently a hot topic in the field of remote sensing research.Currently,existing deep learning remote sensing image semantic segmentation models often have complex network structures and a large number of parameters,and are difficult to apply on mobile devices.Therefore,maintaining a balance between the complexity and segmentation accuracy of remote sensing semantic segmentation models is of great practical significance.This paper conducts in-depth research on the lightweight remote sensing image semantic segmentation model based on DeeplabV3+,and the main work includes the following points:1.To achieve an efficient and accurate semantic segmentation model,Mobile Net V2 is used as the backbone network.However,the main structure of Mobile Net V2 is insufficient for remote sensing image semantic segmentation tasks,and its feature extraction and spatial perception capabilities are limited.Therefore,this paper improves Mobile Net V2 by using the residual structure in Rep VGG to enhance the feature extraction and spatial perception capabilities of the backbone network,making it perform better on remote sensing image semantic segmentation tasks.2.The combination of Dice Loss and Cross-Entropy loss functions is used as the loss function to solve the problem of sample area imbalance in remote sensing images.The Dice Loss function can better focus on the segmentation performance of small area categories,while the Cross-Entropy loss function can ensure the balance of overall categories.This combination can effectively solve the problem of sample area imbalance and improve the model segmentation performance.3.By improving the ASPP module and decoder in DeeplabV3+,the semantic segmentation model parameter and computational complexity are reduced.Firstly,deep separable convolution is used to replace the 3×3 convolution in the ASPP module and decoder to reduce the model parameter and computational complexity and enhance local feature extraction ability.Secondly,3×3 group convolution is used to replace the 3×3 convolution in the backbone network,which can further reduce the model parameter and computational complexity but also reduce a certain degree of model segmentation accuracy.To solve the problem of accuracy loss,ECA attention mechanism is used to enhance channel.By organically combining ECA attention mechanism with group convolution network,the model’s segmentation accuracy and generalization ability are maintained while reducing model computational complexity,parameter quantity and complexity.The semantic segmentation model proposed in this paper was experimentally verified on the WHDLD,DLRSD and UDD6 remote sensing image datasets.The experiments show that the method proposed in this paper has achieved good results and has strong practical value in actual remote sensing image segmentation applications.