An Improved YOLOv7-tiny-based Lightweight Network for the Identification of Fish Species

In order to improve the speed of object detection, an improved YOLOv7-tiny-based lightweight network is proposed in this paper. The lightweight model, named as YOLOv7-tinyMobileOne model, is constructed by introducing a variant of MobileOne to replace the backbone network of YOLOv7-tiny. This model reduces the number of parameters and calculation expenses in the inference phase by pruning the network branch structure. Meanwhile, it alleviates the delay phenomenon of terminal devices. Then the proposed model achieves a marked improvement in detection efficiency, suitable for deployment on terminal devices. To verify the performance of the proposed model, we collected and annotated 7085 images of 17 fish species, and the following conclusions can be drawn from this dataset: (1) compared to YOLOv7, the parameter amount and the computational cost of YOLOv7-tiny-MobileOne decreased by 87.67% and by 89.94%, respectively, while the FPS of YOLOv7tiny-MobileOne increased by 284%; (2) compared to YOLOv7tiny, the parameter amount, and the computational cost of YOLOv7-tiny-MobileOne decreased by 25.45% and by 21.21%, respectively, while the FPS of YOLOv7-tiny-MobileOne increased by 9.94%. The experiments demonstrate that the improved model has a smaller volume and can achieve approximate accuracy while significantly improving identification speed, providing considerations for model deployment of terminal devices.