A Lightweight Tea Bud Detection Model Based on Yolov5

Abstract

Tea bud detection technology is highly significant in realizing the automation and intelligence of tea bud picking. However, there are still some challenges with tea bud detection technology. For example, the problems of low detection accuracy, heavy computing, and large detection model size make the technology inconducive to the deployment of mobile terminals. Thus, a lightweight tea bud detection model based on the Yolov5 model was proposed in this study. Improvements were made in the following aspects: the Ghost_conv module was introduced to replace the original convolution, considerably reducing the computing and model size; the bottleneck attention module (BAM) was added to the backbone network to suppress invalid information and improve the model detection accuracy; the weighted feature fusion was used in the neck network to efficiently fuse the low-level and high-level features, helping the network to extract effective information for recognition and improve detection accuracy; and CIoU was used as the bounding box loss function to accelerate regression prediction and improve the positioning accuracy of the bounding box of the model. A test was conducted on the collected dataset to verify whether the modified model improved the detection performance of tea buds. The results showed that compared with that of the original Yolov5 model, the mean average precision of the modified model increased by 9.66%, and the floating-point operations and params reduced by 52.402 G and 22.71 M, respectively. An ablation study proved that the proposed method improves the detection performance of the Yolov5 model for tea buds. Compared with other detection algorithm models, the superiority of the proposed algorithm in tea bud detection can be seen. The proposed improved Yolov5 can effectively detect tea buds, which provides theoretical research and technical support for intelligent picking of tea buds in actual scenes.

Keywords: Tea bud detection, deep learning, Lightweight, Attention mechanism, Weighted feature fusion