Irregular Defect Detection Based on Adaptive Anchor and Multiscale Feature Fusion

This paper focuses on addressing defect detection in large plastic film production lines, emphasizing model construction and dataset building. Unlike conventional industrial defect detection, these production lines pose a highly complex environment due to high temperatures and strong vibrations. Detecting defects in plastic film is challenging yet crucial for quality control. This research area has received limited attention, and existing methods are inadequate for industrial applications. Plastic film defects are irregular and difficult to characterize, making detection challenging due to the small proportion of defective pixels. Our proposed MRCNet aims to accurately detect these irregular defects using the RaFPN module for multi-scale feature fusion and the Multiple Adaptive Anchor module for versatile anchor frame detection. We also introduce the Contextual Attention module for better detection of small objects in plastic films. MRCNet achieves an [email protected] of 76.8% on our IDD dataset, surpassing current state-of-the-art detectors. Additionally, we present the IDD dataset, curated from multiple production lines, despite considerable time and cost.