{"title":"A Cityscape Image Detail Extraction Enhancement Method for Lightweight Semantic Segmentation","authors":"Xinhe Yu, Huarong Xu, Lifen Weng","doi":"10.1109/ASID56930.2022.9995858","DOIUrl":null,"url":null,"abstract":"Lightweight semantic segmentation is widely used in automotive driving. But the existing methods lack the ability to extract the detailed features of urban street scenes, and the semantic segmentation network structure lacks the logical relationship of interdependence. In order to improve semantic segmentation performance in automotive driving, this paper is based on BisenetV2 to propose: (1) The re-parametrization strategy to improve the ability to extract details features. (2) The SENet channel attention mechanism is adopted to explicitly establish the interdependence between feature channels. (3) Using the larger kernel in the deep layer of the network structure increases the accuracy of semantic segmentation and hardly affects the calculated amount. We tested the Cityscapes test dataset to achieve 72.23% mIoU at 2048×1024 resolution with the speed of 39.55 FPS on one NVIDIA RTX A5000 card without pre-training and accelerated implementations like TensorRT, which is 1.8% more accurate than the latest methods while almost as fast.","PeriodicalId":183908,"journal":{"name":"2022 IEEE 16th International Conference on Anti-counterfeiting, Security, and Identification (ASID)","volume":"23 1","pages":"0"},"PeriodicalIF":0.0000,"publicationDate":"2022-12-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"1","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"2022 IEEE 16th International Conference on Anti-counterfeiting, Security, and Identification (ASID)","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/ASID56930.2022.9995858","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
引用次数: 1
Abstract
Lightweight semantic segmentation is widely used in automotive driving. But the existing methods lack the ability to extract the detailed features of urban street scenes, and the semantic segmentation network structure lacks the logical relationship of interdependence. In order to improve semantic segmentation performance in automotive driving, this paper is based on BisenetV2 to propose: (1) The re-parametrization strategy to improve the ability to extract details features. (2) The SENet channel attention mechanism is adopted to explicitly establish the interdependence between feature channels. (3) Using the larger kernel in the deep layer of the network structure increases the accuracy of semantic segmentation and hardly affects the calculated amount. We tested the Cityscapes test dataset to achieve 72.23% mIoU at 2048×1024 resolution with the speed of 39.55 FPS on one NVIDIA RTX A5000 card without pre-training and accelerated implementations like TensorRT, which is 1.8% more accurate than the latest methods while almost as fast.