YOLO-MTG: a lightweight YOLO model for multi-target garbage detection

With wide adoption of deep learning technology in AI, intelligent garbage detection has become a hot research topic. However, existing datasets currently used for garbage detection rarely involves multi-category and multi-target garbage that are densely accumulated in actual garbage detection scenarios. In addition, many existing garbage detection models have such problems as low detection efficiency and difficulties in integration with resource-constrained devices. To address the above situations, this study proposes a lightweight YOLO model for multi-target garbage detection (YOLO-MTG). This model is designed as follows: firstly, MobileViTv3, a lightweight hybrid network, serves as the feature extraction network to encode global representations, enhancing the model's ability of discriminating dense targets. Secondly, MobileViT block, the feature extraction unit, is optimized with combination of EfficientFormer and dynamic convolution, aiming to enhance the model's feature extraction capability, focusing on essential feature information and reduce the redundancy in useless information. Finally, feature reuse techniques are deployed to reconstruct Neck to minimize the loss of channel information in the feature transmission process, and maintain the strong feature fusion ability of the model. The experimental results on the self-built multi-target garbage (MTG) dataset show that YOLO-MTG achieves 95.4% mean average precision (mAP) with only 3.4 M parameters, and it is superior to other state-of-the-art (SOTA) methods. This work contributes new insights into the field of garbage detection, aiming to advance garbage classification for practical engineering applications.