基于改进YOLOv8-seg的车辆智能定损

Vehicle Intelligent Damage Assessment Based on Improved YOLOv8-seg

摘要: 在中国科技快速发展的大背景下，智能化、自动化逐渐渗透到各个领域。车辆的智能定损研究正在有序推进，传统的定损过程需要进行现场勘察和测量，这一过程繁琐且耗时。而智能定损技术通过分析大量的汽车相关数据和损伤图像，能够在几秒钟内完成对损伤的识别和评估。但目前相关应用泛化性不足，受图片分辨率影响，在小目标及区域分割精确度上仍有较大进步空间。针对这些问题，本文以YOLOv8-seg图像分割算法为基准模型提出了一种精确度高的车辆智能定损模型。通过引入NWD (Normalized Wasserstein Distance)损失函数，强化了模型的小目标识别及分割能力，以及对低分辨率图像的特征提取能力，提升了模型的鲁棒性。提出DLKA (Deformable Large Kernel Attention)可变形大核注意力机制与YOLOv8-seg主干网络C2f模块相融合，实现了在计算复杂度略微增加的前提下，模型分割能力的大幅度提升，强化模型的泛化能力。通过实验验证，本文的改进算法相比原始算法，平均精度上提升了8.1%，在车辆智能定损研究中表现出色。

Abstract: In the context of China’s rapidly developing technology, intelligence and automation are gradually permeating various fields. Research on intelligent vehicle damage assessment is being carried out in an orderly manner. Traditional damage assessment processes require on-site inspections and measurements, which can be cumbersome and time-consuming. Intelligent damage assessment technology analyzes large amounts of car-related data and damage images to identify and evaluate damages within seconds. However, current applications lack generalization capabilities and are affected by image resolution, leaving significant room for improvement in accuracy for small targets and regional segmentation. In response to these issues, this paper proposes a high-precision intelligent vehicle damage assessment model based on the YOLOv8-seg image segmentation algorithm. By introducing the NWD (Normalized Wasserstein Distance) loss function, the model enhances its ability to recognize and segment small targets as well as extract features from low-resolution images, thereby improving its robustness. The proposed DLKA (Deformable Large Kernel Attention) mechanism is integrated with the C2f module of the YOLOv8-seg backbone network, achieving a substantial increase in model segmentation capability with only a slight increase in computational complexity, thus enhancing the model’s generalization ability. Experimental validation shows that the improved algorithm presented in this paper achieves an average precision improvement of 5.9% compared to the original algorithm, demonstrating excellent performance in research on intelligent vehicle damage assessment.