Detection and segmentation of tongue maps containing negative samples based on YOLOv7 and TSRNet

This study investigates the impact of training the TSRNet, a deep convolutional neural network for tongue segmentation, with a dataset comprising both positive (with tongue body) and negative (without tongue body) samples. We introduce an artificial classification mechanism to enhance TSRNet by manually labeling negative samples and integrating them with positive samples. This method significantly improves the network’s ability to distinguish between tongue and non-tongue regions, thereby strengthening its generalization capability. Furthermore, we address the challenge of environmental factors affecting tongue image detection by implementing the advanced YOLOv7 model for more precise localization and detection, enhancing the TSRNet’s accuracy. Validation using an opensource tongue image database demonstrates the efficacy of TSRNet, even when trained on a dataset predominantly consisting of negative samples, in accurately segmenting tongue images.