| Summary: | With the automation of industrial production, appearance defect detection based on machine vision plays an important role in product quality control. The scarcity of defect samples and real-time requirement are the main challenges in this field. Many existing studies are based on semantic segmentation network, but they cannot provide a classification confidence score for each image and only report the segmentation tasks metrics, which ignore that the positive or negative decisions are the key of defect detection. Therefore, this paper proposes a four-stage appearance defect detection model: contrast enhancement, segmentation, correction, and decision, which can achieve high detection accuracy with a severe shortage of positive samples. Since the proposed model simplifies U-Net to segment those candidate defect regions, and constructs a lightweight decision network based on the candidate regions and segmented mask, the proposed method not only achieves fast inference speed, but also obtain good performance with fewer defect samples. Experiments are implemented on three public datasets: magnetic tile dataset, Kolektor surface defect dataset and DAGM2007 dataset. The influence of each module on the detection accuracy is analyzed. Experimental results show that the proposed model achieves excellent performance comparing with other state-of-art methods.
|
|---|