Low contrast stamped dates recognition for pill packaging boxes based on YOLO-SFD and image fusion.

Dates sealed by embossed stamps (stamped dates) on pill packaging boxes are crucial for people to notice the expiration dates. Thus, manufacturers must ensure the accuracy of stamped dates on the pill packaging boxes. However, the low contrast between the stamped dates and the surface of the container and the lack of clear character outlines often lead to light interference, which hampers the stamped dates detections by various deep learning models. This article presents a method for detecting stamped characters that combines YOLO-Shallow Feature Detection (YOLO-SFD) and image fusion. YOLO-SFD enhances YOLOv8 by introducing a Shallow Feedback Feature Pyramid Network (SF-FPN) that includes shallow feature paths and feedback paths for the original input. This enables the model to pay more attention to the shallow features, which include a more detailed character contour. Secondly, this paper presents a lightweight convolutional structure to enhance the model's portability without sacrificing accuracy. Experimental results indicate that SF-FPN outperforms the original feature pyramid network in terms of accuracy and model size. On the stamped date dataset, YOLO-SFD achieves an average accuracy of 99.02 %, 2.57 % higher than the original model, with 29 % fewer parameters. In addition, a 1.5 % mAP improvement is achieved on the Street View House Numbers Dataset. This shows that this method can better fulfill the requirements of pill packaging box stamped date detection and provides new solutions and ideas for related character detection or low contrast image detection. [ABSTRACT FROM AUTHOR]