Car body surface damage detection is a classic problem in computer vision. The main bottleneck of car body surface damage detection lies in the different scales of damage instances in the image， which affects the accuracy and efficiency of segmentation. In this paper， we use a single-stage semantic segmentation network （YOLACT⁺⁺） for damage detection on the car body surface， combine EfficientNet to design a backbone network to improve segmentation efficiency， and improve the loss function optimization YOLACT⁺⁺ to generate the target instance Mask in the experiment. Experimental data are marked by deep learning， and results show that the improved YOLACT⁺⁺ detection frame rate is increased to 35 frame/s， which reduces the mask generation error and improves the instance segmentation accuracy of YOLACT⁺⁺.