Transformer-Based End-to-End Vehicle Leather Defect Detection Method

Currently, the manual inspection and marking of vehicle leather defects heavily relies on the subjective judgment of personnel, leading to significant error margins and requiring highly qualified personnel, resulting in an inability to guarantee the accuracy of test results. To address these issues, this research proposes a Transformer-based end-to-end algorithm for detecting vehicle leather defects, named D2TR (Defect Detection Transformer). The D2TR can improve the efficiency and accuracy of leather defect detection, save on manpower, reduce production costs, and enhance the intelligence of leather processing, When compared to manual detection. Firstly, a Convolutional Neural Network (CNN) layer is employed to extract leather defect features. Then the features are fed into a Transformer-based encoding-decoder architecture to generate a set of target defect prediction boxes, which are used to predict the detection results through a Feedforward Neural Network (FFN). Additionally, data augmentation method is used to expand the sample size. The average detection accuracy of D2TR on our self-collected leather defect dataset was found to be 81.2%.