基于改进YOLOv7的湖面漂浮物目标检测算法OACSTPCD

This research strives to improve the accuracy and reasoning speed of detection and identification of multi-species and small volume floating garbage on the lake.In combination with the image characteristics of floating garbage on the lake,the X-Toss pruning framework is established by adopting semi-structured pruning technology.On the basis of object detection model YOLOv7,a lightweight real-time detection method C-X-YOLOv7 for floating objects on the lake is proposed.In the X-Toss pruning framework,the DFS(depth first search)algorithm is used to generate parent-child convolution kernel calculation graph,and specific kernel mode is used to prune convolution kernel to reduce the computational cost of iterative pruning.The model is weighted by combining CA(coordinate attention)mechanism to reduce the over-fitting of the model and improve its accuracy and generalization ability.The results show that the recognition accuracy of the model C-X-YOLOv7 is 91.7％and its recall rate is 91.2％,which is 2.6％ and 2.5％ higher than those of the model YOLOv7,respectively.In terms of the inference acceleration,the X-Toss pruning framework achieves the acceleration ratio of 1.98×and 2.17×of YOLOv7 on RTX 2080 Ti and NVIDIA Jetson TX2,respectively.The acceleration ratio and energy consumption of X-Toss are improved in comparison with those of the pruning frameworks such as PD,NMS and NS.The research shows that the floating object detection method C-X-YOLOv7 can provide a new idea for the detection and identification of lake garbage.