A drug identification model developed using instance segmentation

Medicine boxes recognition is an important process in several sectors and industries, as pharmacies and the pharmacy industry or hospitals. A single failure in this procedure can have consequences that affect both on a human level and on a large-scale legal level. Due to these factors, it is necessary to address the problem of how to identify medicine packages based on deep learning techniques, in concrete, using the Convolutional Neural Network. However, several of the algorithms previously applied have failed to obtain accurate results when objects or logos were too small or the scene was not completely clear. The paper has been focused on the improvement of efficiency in object detection and recognition algorithms based on Mask detection. Also, it has been covered the need to find a dataset based on small objects and pharmacist products. The model is implemented with COCO dataset and a custom dataset of medicine packages. The evaluation carried out on a cloud platform, compares the algorithm Yolov4 and Mask R-CNN combined with the backbone ResNet50, as a result, Mask R-CNN needs less time for training and has higher performance. The effectiveness raised for the mask segmentation architecture was higher than 95%, in consequence, it is an acceptable score for the real world