Intelligent Defect Detection System in Pharmaceutical Blisters Using YOLOv7

Quality assurance of blister packages for defect detection is necessary, but manual inspection is inefficient. The proposed methodology develops an automated defect detection system using the YOLOv7 to detect five defect categories: broken pill, crack pill, empty pill, foreign object, and color mismatch. The proposed model is initially trained for 20 epochs without fine-tuning the hyperparameters, then empirically, the epochs and momentum were fine-tuned to observe the performance improvement. The fine-tuning approach improved precision and recall by 6% and 5% of the proposed YOLOv7, respectively. The model also achieves high accuracy with a mAP@0.5 score of 0.962. A web interface is developed which allows users to upload images to identify the occurrence of defects in blister packages. The proposed methodology has developed blister package defect detection, significantly enhancing the quality assurance of blister packaging by improving efficiency and accuracy.