Research On Semantic Segmentation Based On Lightweight High Resolution Network

Semantic segmentation is one of the four basic tasks in the field of computer vision,and has a wide range of application scenarios in the fields of automatic driving,medical image recognition,and intelligent transportation.The task of semantic segmentation is to predict a predefined class label for each pixel in an image.Most of the existing semantic segmentation results are based on deep learning technology.However,deep learning technology cannot be well deployed on mobile platforms with limited computing resources.Performance has become one of the research hotspots in current image semantic segmentation.This paper aims at the problems of inaccurate segmentation of object boundaries and weak contextual relevance of semantic category objects in most networks,a real-time semantic segmentation algorithm based on boundary-aware is proposed,and a semantic segmentation algorithm based on dense recurrent structure and Transformer,and the efficiency of the proposed algorithm is verified on the semantic segmentation dataset,the main research contents and innovations are as follows:(1)Real-time semantic segmentation algorithm based on Boundary-Aware.Aiming at the problem of inaccurate object boundary recognition in the existing semantic segmentation model,which leads to the decline of the model segmentation accuracy,the boundary-aware learning mechanism is first introduced into the lightweight high-resolution Bi Se Net V2,and the horizontal and vertical position information is added to the detail branch and detail boundary features.Second,a lightweight region adaptive module is designed to enhance the ability of convolutional neural networks to model the complexity of semantic boundary regions.Finally,an distinctive atrous spatial pyramid pooling module is designed to further highlight details and semantic features according to the different contribution values of pixels in the sampling area.Ablation and comparison experiments on Cityscapes and Cam Vid datasets show that the proposed algorithm can effectively improve the segmentation quality of object boundaries.(2)Semantic Segmentation Algorithm Based on Dense Recurrent Structure and Transformer.Aiming at the problem of weak contextual relevance of semantic category objects,the aggregation module of Bi Se Net V2 network is replaced with an integrated dense loop module,and a semantic segmentation algorithm based on dense recurrent structure and Transformer is proposed.First,a dense recurrent structure is employed to effectively enhance the association of semantic objects by providing historical information for the feature information output by adjacent convolutions.Secondly,the inner-outer Transformer modules are designed,and the semantic information is given appropriate weights by adopting the design of the inner and outer layers of self-attention mechanism.Finally,a shared-guided fusion module is designed,which fuses the semantic object information of the semantic and detail branches to further improve the communication efficiency between the two branches,making the method of dual-branch fusion information more efficient.Experiments show that this algorithm can improve the quality of semantic segmentation network by enhancing the relevance of semantic objects.There are 21 pictures,14 tables,and 66 references in this thesis.