Performance Review of Modern AI Algorithms Utilized for Medical Waste Sorting Works

This paper evaluates the performance of modern AI-based object detection models that can be used for object classification and sorting applications. In this case, we focused on the classification of the medical waste for the current global situation which is the medical waste management during the post-pandemic of Covid-19 phase. A few classification models were used and compared between (1) CNN and ResNet50 and (2) YOLO v3 and YOLO v4. The results also were compared with the previous works that focused on waste classification. The difference between this work is the image dataset, which our work train and test the medical waste (facemask, glove, and syringe), while the previous works focused on general waste such as food, plastic, metal, paper, and others. From 2207 images of the medical waste, CNN and ResNet achieved 89.35 and 85.75% of accuracy, respectively, where it requires more images per class for the training improvement. YOLO v3 and YOLO v4 used 3073 images for training and achieved 84.86 and 89.21% of mean average precision (mAP). Our YOLO v3 mAP is in the average value among the previous works, while YOLO v4 has a higher mAP compared to the YOLO v4 training from other works. The YOLO v4 then was tested in real-time medical waste detection and successfully detected the masks, gloves, and syringe. However, there are still some wrong detections during the real-time detection using the camera, especially with other objects with similar shapes to the medical waste. Further, performance evaluations are required that can be used for medical waste objects and also for other different objects based on the applications.