Special-shaped cigarette detection algorithm based on improved YOLOv5

In order to solve the problem of heavy workload in checking and detecting special-shaped cigarettes, reduce the false detection rate, and improve the working efficiency and intelligent level of the tobacco logistics industry, a special-shaped cigarette detection algorithm based on improved YOLOv5 is proposed. The algorithm takes YOLOv5 as the basic framework, introduces pyramid split attention(PSA) module to extract feature space domain information, and concatenate with feature channel domain information extracted by Focus module to obtain multi-dimensional feature information; Atrous Spatial Pyramid Pooling (ASPP) is used to replace SPP module, and use different ratio of cavity convolution operation to expand receptive field and reduce operation parameters while ensuring resolution; K-means++ algorithm is used to optimize the anchor frame and improve the matching degree between the anchor frame and the obtained image. The algorithm was tested with the special-shaped cigarette image set collected in the production line, and the mAP parameter of the experimental result is 94.14%, which is 5.04% higher than the result before optimization. In addition, the improved model does not introduce complex modules, and can be deployed in the AI edge computing integrated development board. The reasoning and recognition time is about 75 ms, which can meet the real-time requirements of the production line.