基于改进YOLOv5的舌面特征检测

Tongue Feature Detection Based on Improved YOLOv5

摘要:

针对传统中医舌诊视觉诊断存在主观性强且耗费精力的问题，提出一种基于改进YOLOv5的舌面齿痕和裂纹特征自动检测模型：该模型在YOLOv5模型的骨干网络中引入SimAM-CSP模块以增强网络的特征提取能力，在瓶颈层和预测部分之间加入瓶颈注意力模块，进一步聚焦关键信息；通过调整YOLOv5的特征融合结构，增加对图像细节的感知能力，提高网络性能；将定位损失函数GIoU替换为EIoU，提升模型的训练收敛速度和预测回归精度；利用K-Means聚类算法对YOLOv5的初始锚框进行调整，使算法更加契合舌面齿痕和裂纹特征检测；将改进后的YOLOv5模型在自制舌象数据集中进行训练，得到的平均检测精度（mAP）为79.5%，较原算法提升了6.3个百分点。实验结果表明改进YOLOv5模型能够有效提高舌面齿痕和裂纹特征检测精度，有助于辅助医生诊断。

Abstract:

Considering the problems of subjectivity and energy-consuming visual diagnosis of traditional Chinese medicine tongue diagnosis, a detection model based improved YOLOv5 for tongue tooth mark and fissure features was proposed. The SimAM-CSP module was introduced into the backbone of YOLOv5 to enhance the feature extraction capability of the network ground. The Bottleneck Attention Module was added between the Neck layer and the Head layer to further focus critical information. The feature fusion structure of YOLOv5 was adapted to increase the ability to perceive image details and improve performance of network. The localization loss function GIoU was replaced with EIoU in original YOLOv5 algorithm for the purpose of simultaneously improving the training convergence speed and prediction regression accuracy. The initial anchor frames of YOLOv5 were adjusted by the K-Means algorithm to make the model more suitable for tongue tooth mark and fissure detection. The improved YOLOv5 model was trained in the self-built tongue image dataset, and the average detection accuracy reaches 79.5%, which is 6.3 percentage points higher than that of the original algorithm. Experimental results show that the improved YOLOv5 model can effectively improve the detection accuracy of tongue tooth mark and fissure, which is helpful for assisting doctors in diagnosis.