Yongsheng Dong , Chongchong Mao , Lintao Zheng , Qingtao Wu , Mingchuan Zhang , Xuelong Li
{"title":"AFPN: Alignment feature pyramid network for real-time semantic segmentation","authors":"Yongsheng Dong , Chongchong Mao , Lintao Zheng , Qingtao Wu , Mingchuan Zhang , Xuelong Li","doi":"10.1016/j.patcog.2025.112019","DOIUrl":null,"url":null,"abstract":"<div><div>The structures of two pathways and the feature pyramid network (FPN) have achieved advanced performance in semantic segmentation. These two types of structures adopt different approaches to fuse low-level (shallow layer) spatial information and high-level (deep layer) semantic information. However, the segmentation results still lack local details due to the loss of information caused by simply fusing low-level feature details directly with multi-level deep features. To alleviate this problem, we propose an alignment feature pyramid network (AFPN) for real-time semantic segmentation. It can efficiently utilize both the low-level spatial information and high-level semantic information. Specifically, our AFPN consists of two components: the pooling enhancement attention block (PEAB) and the dual pooling alignment block (DPAB). The PEAB can effectively extract global information by using an aggregation pooling operation. The DPAB performs two types of pooling operations along the channel and spatial dimensions, reducing the differences between multi-scale feature maps. Extensive experiments show that AFPN achieves a better balance between accuracy and speed. On the Cityscapes, CamVid, and ADE20K datasets, AFPN achieves 78.75%, 79.24%, and 39.56% mIoU and the speed meets the real-time requirement. Our code can be available at the <span><span>https://github.com/chongchongmao/AFPN</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49713,"journal":{"name":"Pattern Recognition","volume":"170 ","pages":"Article 112019"},"PeriodicalIF":7.5000,"publicationDate":"2025-06-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Pattern Recognition","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S003132032500679X","RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
The structures of two pathways and the feature pyramid network (FPN) have achieved advanced performance in semantic segmentation. These two types of structures adopt different approaches to fuse low-level (shallow layer) spatial information and high-level (deep layer) semantic information. However, the segmentation results still lack local details due to the loss of information caused by simply fusing low-level feature details directly with multi-level deep features. To alleviate this problem, we propose an alignment feature pyramid network (AFPN) for real-time semantic segmentation. It can efficiently utilize both the low-level spatial information and high-level semantic information. Specifically, our AFPN consists of two components: the pooling enhancement attention block (PEAB) and the dual pooling alignment block (DPAB). The PEAB can effectively extract global information by using an aggregation pooling operation. The DPAB performs two types of pooling operations along the channel and spatial dimensions, reducing the differences between multi-scale feature maps. Extensive experiments show that AFPN achieves a better balance between accuracy and speed. On the Cityscapes, CamVid, and ADE20K datasets, AFPN achieves 78.75%, 79.24%, and 39.56% mIoU and the speed meets the real-time requirement. Our code can be available at the https://github.com/chongchongmao/AFPN.
期刊介绍:
The field of Pattern Recognition is both mature and rapidly evolving, playing a crucial role in various related fields such as computer vision, image processing, text analysis, and neural networks. It closely intersects with machine learning and is being applied in emerging areas like biometrics, bioinformatics, multimedia data analysis, and data science. The journal Pattern Recognition, established half a century ago during the early days of computer science, has since grown significantly in scope and influence.