Small defect detection in panel display industry by faster region convolutional neural network with attentional feature fusion

Detecting small defects in the presence of complex background interference remains a significant challenge in automated optical inspection (AOI) for quality control. Traditional threshold-based methods struggle to identify these defects due to their small size relative to the overall image. This study aims to enhance defect detection performance by employing an Attentional Feature Fusion (AFF) approach. To address this challenge and improve small defect detection accuracy, this study proposes an enhanced Faster Region-Based Convolutional Neural Network (Faster-RCNN) with Attentional Feature Fusion (FRCNN_AFFNet). The AFF mechanism dynamically adjusts feature weighting, effectively integrating low-level and high-level features to refine defect identification. To validate the proposed method, an empirical study was conducted in panel display manufacturing, where over 90% of the defects were classified as small-sized. Experimental results show that FRCNN_AFFNet improves mean average precision (mAP) by 6.41% compared to standard Faster-RCNN. The AFF module enhances the performance of Region Proposal Networks (RPNs) by optimizing feature selection, leading to more accurate bounding box proposals. These findings demonstrate the practical effectiveness of FRCNN_AFFNet in improving small defect detection in industrial applications.