FGNet: a lightweight network for tongue image segmentation

In contemporary medical technologies, tongue imaging plays a crucial role in diagnosing various diseases, particularly within traditional Chinese medicine, where tongue diagnosis is a critical component. This study aims to develop a lightweight and rapid semantic segmentation network tailored for tongue images. We highlight the importance of low-level features for precise pixel-level predictions. Additionally, we explore the impact of arbitrary histogram equalization on network performance degradation during image enhancement. Experimental results demonstrate that our model, which utilizes only 3.8% of the parameters compared to UNet and requires approximately 81.44%of the computation time, achieves a prediction mean Intersection over Union (MIoU) that is 5.49% higher. This indicates that our model is well-suited for deployment on compact embedded systems.