Research on the improvement of garbage detection and classification algorithm based on YOLOv7-tiny

This study introduces an optimized version of the YOLOv7-tiny algorithm, enhanced to boost its efficiency and effectiveness in garbage detection and classification. Key innovations include a new detection head and an updated loss function within a feature fusion network. Firstly, an improved lightweight C3Shufflev2 structure based on ShuffleNetv2 is proposed, replacing the complex residual network modules in the original network to reduce the number of network parameters. Secondly, the efficient RepBiPAN structure is used in the feature fusion network part to enhance the network’s feature fusion ability without adding extra network parameters. Then, a new AsDDet is used in the detection head part, increasing the network’s focus on small targets and reducing the miss rate. Finally, the Inner-IoU optimization algorithm’s coordinate loss function is used to avoid instability in the training process caused by the failure of the penalty function, thereby enhancing the network’s robustness. Performance assessments were conducted using a proprietary Garbage dataset, revealing that this refined model outperforms its predecessor by enhancing accuracy by $\mathbf{2 . 7 \%}$, while significantly reducing both the floating-point operations by $46.6 \%$ and the model size by $26.5 \%$.