Improved ConvNeXt Integrating Efficient Multi-Scale Attention and Focal Loss for Vehicle Exterior Damage Recognition

For the past few years, computer vision technology has been widely used in practical problems. Currently, the damage determination process in the field of automobile claims requires a large amount of manpower, and the instance segmentation technique based on convolutional neural networks can greatly reduce the cost of the industry. For the mission of vehicle appearance damage recognition, we compared the performance of various algorithms on CarDD, among which the ConvNeXt algorithm performed well. Based on this, the efficient multi-scale attention (EMA) is introduced into the backbone network to enhance its feature extraction capability. Meanwhile, for shape-variable type damage, focal loss was used to help the model focus on difficult categories. Finally, we proposed an algorithm named ConvNeXtFA, which significantly improved the overall performance of vehicle damage recognition and effectively improved the detection of difficult category vehicle damage.