针对车辆智能定损领域中存在的损伤目标类间特征相似性高而导致的区分效果较差等问题，提出了一种基于改进YOLOv7的车辆变形类检测方法（vehicle deformation detection, VDD-YOLO）。首先，在模型的主干网络加入空间的混合注意力模块（mixed spatial channel attention, MSCA），使模型能够提取更加丰富的样本特征，进而提升对判别性特征的提取能力；其次，对损失函数进行优化，使用Focal-EIoU作为损失函数增强模型锚框回归质量，进而提高模型定位精确率；最后，针对变形损伤中的难检测样本，使用了一种平衡因子调控方法，使得模型更专注于难检测样本的特征捕捉，从而提高模型对难检测类别的识别效果。实验结果表明，VDD-YOLO算法提高了三类损伤的平均检测精度，mAP50达到43.9%，相较于基线模型提高3.8%。此外，所提出的算法相较于其他主流先进实例分割算法在模型精度和困难样本检测上均有显著优越性，为车辆智能定损提供了一种新的解决方案。

In order to solve the problem of poor discrimination performance caused by high similarity between damage targets in the field of vehicle intelligent damage determination, a vehicle deformation detection method based on improved YOLOv7 was proposed. Firstly, mixed spatial channel attention (MSCA) was added to the backbone network of the model to enable the model to extract more abundant sample features, and thus improve the ability to extract discriminative features. Secondly, the loss function was optimized, and Focal-EIoU was used as a loss function to enhance the quality of anchor frame regression and improve the accuracy of model positioning. Finally, a balance factor regulation method was used for the hard-to-detect samples in deformation damage, which makes the model focus more on the feature capture of hard-to-detect samples, so as to improve the recognition effect of the model on the hard-to-detect categories. The experimental results show that the VDD-YOLO algorithm can improve the average detection accuracy of three types of damage, and the mAP50 reaches 43.9%, which is 3.8% higher than that of the baseline model. In addition, compared with other mainstream advanced case segmentation algorithms, the proposed algorithm has significant advantages in model accuracy and difficult sample detection, and provides a new solution for vehicle intelligent loss determination.