Garbage Object Detection Based on Improved YOLOv5

Abstract

Traditional waste classification methods often rely on manual feature extraction and manual rule design, which cannot meet the needs of practical application scenarios. This thesis proposes a YOLOv5^[1] target detection method for garbage classification, aiming to solve the problem of automation and efficiency in garbage classification. First, we introduce semi-supervision to enhance the model's detection robustness as well as its extensiveness. Second, we incorporate Mixup, which fuses multiple different datasets to obtain new images with occlusions, which are trained in order to improve the robustness of the model for targets with complex backgrounds. In addition, we introduce a mosaic data enhancement algorithm to defocus the target and improve the accuracy of target detection. The proposed garbage target detection method based on the improved YOLOv5 has made significant progress in improving the detection performance, providing more effective tools and technical support for environmental protection and urban management. This method is expected to be widely popularized in practical applications.