关键词: 茶叶嫩芽识别, AD-YOLOX-Nano算法, 注意力机制, 深度可分离卷积

Abstract: In order to address the difficulties in identifying tea buds and improve the accuracy and robustness of tea bud recognition in natural environments, this paper develops an improved YOLOX-Nano algorithm （AD-YOLOX-Nano） by integrating attention mechanism and depthwise separable convolution. The algorithm is based on the YOLOX-Nano model, which uses CSPDarkNet as the backbone network. It reduces the workload of feature extraction by introducing Depthwise Separable Convolution in the CSPDarkNet network. The Convolutional Block Attention Module is incorporated into the feature pyramid of the YOLOX-Nano network to learn the feature correlation of different channels, enhance the transmission of depth information in the network, and improve the recognition capability of tea buds in different scenarios. The experimental results show that the AP value and F1 values of the AD-YOLOX-Nano algorithm are 85.6% and 86%, respectively. Compared to the YOLOX-Nano algorithm in the same environment, this algorithm maintains a similar model size, but achieves an improvement of 2.7% in AP value and 3% in F1 value. Compared to the commonly used object detection algorithms such as YOLOv5-S, YOLOv4, and Faster R-CNN, the AD-YOLOX-Nano algorithm has a model size of only 1/7 of theirs. But it achieves an improvement of 5.4%, 5.5%, and 6.28% in AP value, respectively. The proposed algorithm has significant advantages in terms of model lightweighting and detection accuracy, providing an effective solution for the deployment of embedded hardware for intelligent tea picking.

Key words:  tea buds recognition, AD-YOLOX-Nano, attention mechanism, depthwise separable convolution