Integrating Bottleneck Attention Module (BAM) into YOLOv8 for Automated Industrial Label Quality Inspection

Abstract

This study proposes an automated inspection solution to address the vulnerability of manual Quality Control (QC) systems to human error in detecting expired date labeling defects on plastic bottle products used for liquid packaging. The developed system employs the single-stage YOLOv8 architecture for object detection, which offers high inference speed, a crucial aspect for real-time applications. This study enhances model accuracy through the integration of a Bottleneck Attention Module (BAM) as an attention mechanism, strategically placed at the 9th layer of the network backbone. The selection of BAM is based on its capability to simultaneously capture channel dependencies and spatial relationships, which is essential for accurately recognizing subtle printing defect patterns in small-sized text. As a result, the enhanced model (YOLOv8-BAM) demonstrates a significant improvement in key performance metrics compared to the baseline YOLOv8 model, namely: an increase in Recall of 4.17%, Precision of 3.23%, and F1-score of 3.16%. These findings validate that YOLOv8-BAM is a more robust and reliable solution for automated industrial label quality inspection.