Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (14): 197-208.DOI: 10.3778/j.issn.1002-8331.2403-0087

• Graphics and Image Processing • Previous Articles     Next Articles

Enhancing YOLOv8 for Improved Instance Segmentation of Automotive Surface Damage

TAN Xu,  ZHAO Ji   

  1. School of Computer Science and Software Engineering, University of Science and Technology Liaoning, Anshan, Liaoning 114051, China
  • Online:2024-07-15 Published:2024-07-15

改进YOLOv8的汽车表面伤损实例分割模型

谭旭,赵骥   

  1. 辽宁科技大学 计算机与软件工程学院,辽宁 鞍山 114051

Abstract: To address the shortcomings of manual damage assessment and issues with conventional vehicle damage detection models in the context of intelligent vehicles, it proposes EIS-YOLO, an enhanced instance segmentation model based on YOLOv8. It introduces CRDB, a novel multi-scale feature fusion and channel reduction module that replaces C2f, reducing parameters by 20.15% while improving fusion efficiency. Additionally, HRFPN structure maintains high-resolution branches, facilitates finer detail and semantic exchange, and includes AFF and BiAM attention modules for deeper feature integration. An efficient E-FPN and an extra output head are utilized to better identify small damages and edges. Evaluated on CarDD dataset, CRDB improves multi-task accuracy by 2?percentage points, and the integrated EIS-YOLO model with HRFPN sees a 4.4?percentage points boost in [PB] and 6.6?percentage points in [PM] over the baseline, all while maintaining a lighter weight and lower computational complexity.

Key words: vehicle damage detection, YOLO-Seg, attention mechanism, multi-scale feature fusion, CarDD vehicle damage data

摘要: 针对人工定损方式无法满足智能汽车时代的发展要求,及传统汽车伤损检测模型精度低、信息少、难部署等问题,提出了改进YOLOv8的汽车伤损实例分割模型EIS-YOLO。在主干网络中设计了一个多尺度特征融合与通道数减小的CRDB模块,取代传统C2f模块,显著减少了参数量的同时提高了特征融合的能力;提出了保留高分辨率分支的HRFPN结构,以加强细节信息保留能力,增强细节与语义信息的交换,该结构通过AFF和BiAM注意力融合模块增强了深层传递,经由简化冗余连接的E-FPN完成特征融合。还增加了一个额外的输出头捕捉细小伤损,提高了模型对小目标伤损及伤损边缘的精确识别。在CarDD数据集上,主干网络部分提出的CRDB模块对比C2f模块实现了同架构下计算量减小20.15%,同时多任务平均准确率提升2个百分点,在此基础上,结合HRFPN结构与额外输出头设计的模型整体的准确率[PB]、[PM]相较于基准模型分别提升了4.4和6.6个百分点,且模型更轻量,计算复杂度更低。

关键词: 汽车伤损检测, YOLO-Seg, 注意力机制, 多尺度特征融合, CarDD汽车伤损数据