Improved YOLOv5 with Backbone Replacement to MobileNet V3s for School Attribute Detection

Abstract

School attributes are a series of clothes and accessories that must be worn by students in the school environment. The implementation of this rule aims to create discipline in students. However, in practice, not all rules can be implemented properly because there are still students who violate these rules. One of the rules applied at school is the use of complete attributes. Currently, attribute checks in schools are done manually or through teacher supervision. However, this takes more time, is prone to errors, and is inefficient due to the large number of students being checked. This study proposes an improved YOLOv5 architecture with the replacement of the backbone to MobileNetV3s to detect school attributes. This method uses deep learning and the YOLOv5 algorithm to detect in real time the use of school attributes by students. In this study, the experimental results show that the enhanced YOLOv5 with MobileNetV3s has higher accuracy compared to the original YOLOv5. In addition, the improved model is more efficient in memory usage and weight file size. With an accuracy result of 0.912 on mAP50 and a weight size of about 90 MB and a memory usage of <7 GB, it shows the potential of replacing the backbone in this technology in overcoming attribute detection challenges in schools and can be applied in other cases. However, further research is needed to generalize these results to other problems. This research also shows that backbone replacement in YOLOv5 can affect the accuracy of the model.