Exploring the Performance Improvement Method of a YOLO-based Mask Appearance Detection Model under Small Data Sets

COVID-19 emerged at the end of 2019 and rapidly affected countries around the globe. In an effort to curb the spread of the virus, many nations implemented various measures, one of which was the widespread use of medical masks. Masks, being more readily available, easier to wear, and offering effective protection compared to goggles and protective clothing, have seen a dramatic increase in usage.In order to facilitate rapid preliminary inspection of mask appearance—specifically the completeness of the mask front and ear loops—on high-volume production lines, this study employs the YOLO object detection framework for mask appearance detection. However, the full YOLOv7 model, with its large architecture and high GPU memory consumption (up to 5.79 GB in our experiments), proves inconvenient for training on less advanced hardware, especially when the task at hand is relatively simple. Thus, YOLOv7-tiny was adopted as the baseline to alleviate computational demands (reducing maximum GPU memory usage to 1.46 GB). Nevertheless, training with YOLOv7-tiny revealed issues with insufficient accuracy. For instance, when the training dataset ranged from 160 to 750 images, the test accuracy varied between 28% and 95%. Consequently, modifications were made to the internal architecture of YOLOv7-tiny—such as adding convolution layers, pooling layers, and implementing cross-layer interactions—to improve accuracy on small datasets while maintaining comparable training, testing, and validation speeds. Under the same dataset conditions (with 160 to 750 training images), the proposed method increased test accuracy by 5% to 32%, achieving overall accuracy levels between 37% and 99%.