CNCR-YOLO: A Comprehensive Optimization Strategy for Small-Target Defect Detection in Injection-Molded Parts
Fei Jiang, Wenjing Ye, Peicong Lu, Shaohui Zhang, Zhaoqian Wu, Haifei Zhu
Abstract
Defect detection in injection-molded parts is crucial for quality control in industrial manufacturing. However, it poses challenges due to process complexity, particularly for detecting small defects. To address this challenge, an efficient detection algorithm named CNCR-you only look once (YOLO) is proposed. First, to both improve the quality and quantity of injection-molded parts defect samples, Style-GANv3 data generation and traditional augmentation method are utilized as preprocessing. Moreover, inspired by YOLOv5, CNCR-YOLO combines the strength of normalized Wasserstein distance and complete intersection over union (IoU) metrics together to raise the CI-normalized Gaussian Wasserstein Distance (NWD) loss function for better optimization. Finally, the convolutional block attention module (CBAM) attention mechanism and the Res2net module are embedded into YOLOv5 to enhance the ability of target feature learning and multiscale feature extraction. The experimental results demonstrate that the proposed method outperforms the other algorithms, achieving an improvement of 1.5% in mean average precision (MAP), notably 2.3% for small-target defects. In addition, its generality and robustness are further validated on an industrial printed circuit board (PCB) defect detection dataset.