The work aims to propose a machine vision-based recognition method for pill boxes with low contrast between the steel-stamped date and the background, inconspicuous character outlines, and recognition susceptible to interference by ambient light. An improved YOLOv5s model was used to correct the collected pill box dataset by perspective transformation and data enhancement. By fusing the Coordinate Attention (CA) in the backbone network of the model, the interference of redundant information was reduced. The neck network introduced weights according to the Bi-directional Feature Pyramid Network (BiFPN) to better balance the feature information of the layers of different sizes. The Wise-IoU (WIoU) was introduced to reduce the intervention of high-quality samples in the training and to improve the model's generalization ability. The experimental results on the self-constructed steel-stamped character dataset showed that the average accuracy of the improved network for recognizing the steel-stamped date of the pill box reached 99.41 %, which was 2.38 % higher than that of the original model, and the frame rate was 80.01 f/s. The improved YOLOv5 model can detect the steel-stamped date of the pill box with a better accuracy than that of the original network, and it can meet the real-time requirement of the production line of the pill box.