Lightweight Oracle Bone Character Detection Algorithm Based on Improved YOLOv7-tiny

Aiming at the problem of difficult recognition caused by the varying scale of oracle characters and the small size of some targets, as well as in order to meet the deployment requirements of application scenarios, a lightweight oracle character detection algorithm based on the improved YOLOv7tiny is proposed. First, Partial Convolution are fused in the model backbone network to reduce the redundant computation and memory footprint of the network model. Second, Asymptotic Feature Pyramid Network (AFPN) is constructed to reduce the problem of detail information loss caused when feature fusion is performed between multiple levels, in order to better capture the features of targets at different scales and enhance the detection of small targets, and reduce model complexity. Finally, a feature fusion network based on the bottleneck residual module is constructed to further reduce the model size and enhance the model deployability, as well as to help the network fuse feature information more efficiently. The experimental results show that the improved model achieved an mAP@0.5 of 90.3%, the number of parameters, computation and model size are reduced by 55.7%, 44.1% and 52.5%, respectively, compared with the base model, and by 75.7%, 74.1% and 74.2% compared to YOLOv8s, respectively. The improved model has been greatly lightweighted and balanced with high accuracy.