Steering knuckle surface defect detection and segmentation based on reverse residual distillation

Although the supervised deep learning method effectively detects and segments the surface defects of the steering knuckle, in the absence of sufficient defect samples, the model is prone to tend to learn normal sample features and ignore defective features during the training process, leading to a higher defect detection error rate. In this paper, we propose an unsupervised defect detection method called reverse residual distillation, which can be trained using only defect-free steering knuckle surface images and can accurately detect and segment surface defects in steering knuckles. In this method, we adopt the encoder-decoder structure as the basic structure of the teacher-student network and integrate the reverse distillation and progressive distillation methods into the distillation process, which solves the overgeneralization problem in the student network and improves the distillation efficiency. Additionally, we introduce a trainable one-class bottleneck embedding module and a multi-scale channel attention feature fusion module to enhance the model's performance in detecting and segmenting defects. Experimental results on the mvtec anomaly detection (MVTec AD) dataset and the steering knuckle dataset demonstrate the effectiveness of our method in detecting and segmenting surface defects in industrial products. Especially in the steering knuckle dataset, the area under the receiver operating characteristic curve (AUROC) scores for defect detection and pixel-level segmentation achieved remarkable levels of 98.6% and 99.8%, respectively.