Cracks-YOLO: an improved YOLOv5 tongue crack detection method based on switchable atrous convolution and transformer

Tongue diagnosis is a vital component of Traditional Chinese Medicine (TCM), with cracked tongue serving as a key diagnostic indicator. Although deep learning techniques have advanced TCM tongue image recognition, existing methods for cracked tongue detection still face challenges of low accuracy and slow processing speed.

To address these issues, we propose an improved YOLOv5 framework integrating Switchable Atrous Convolution (SAC) and a Transformer encoder. First, given the varied morphology, size, and distribution of tongue cracks, SAC is introduced into the backbone network to adaptively adjust receptive fields via different atrous rates, enabling multi-scale feature extraction and better preservation of fine details. Second, a Transformer encoder module is incorporated to capture global information and long-range dependencies, thereby enhancing the network's overall feature representation capability. Experiments are conducted on a self-built tongue image dataset to evaluate the proposed modifications.

The improved YOLOv5 demonstrates superior performance over the original YOLOv5 in tongue crack detection, achieving an increase of 2.2% in AP@50 and 8.43% in AP@50:95 metrics on the self-built dataset.

These results confirm that the combination of SAC and Transformer effectively boosts detection accuracy and generalizability, addressing the limitations of prior approaches. The proposed method offers a more reliable and efficient solution for automated tongue crack analysis, with potential implications for clinical TCM diagnostic support.