Abstract:This study aims to develop an automated method for auricular keypoint detection to compensate for the drawbacks of manual auricular acupoint positioning in traditional Chinese medicine (TCM), such as low efficiency and strong subjectivity, and to promote the objectification and intelligentization of TCM constitution identification. It proposes a convolutional method named group-depth convolution (GDConv). By combining group convolution with depth-separable convolution and then performing a channel shuffle operation, this can reduce the computational cost while maintaining the accuracy to the greatest extent. Meanwhile, the CARAFE module, which is an up-sampling operator, is adopted to increase the receptive field of the network, better restore the details of auricular acupoints, and improve the model’s acupoint detection ability. The experimental results show that compared with the baseline model, the YOLOv8-ear model demonstrates significant lightweight advantages. Its parameter amount is approximately 52.09% lower than that of YOLOv8n, and mAP@0.5 reaches 99.1%. The deep learning method can effectively identify the features of auricular keypoints, providing a new approach for the modernization and intelligent classification of TCM based on real-time auricular acupoint detection.