A novel tongue segmentation method based on improved U-Net

Abstract

Background and objectiveAccurate segmentation of tongue image is a prerequisite to intelligent tongue diagnosis. Current tongue segmentation algorithms have good performances in tongue image segmentation in a standard environment. However, tongue segmentation is still challenging to be used in an open environment due to factors such as the complex external open environment and relatively simple equipment. This study aims to construct a novel deep neural network tongue segmentation system suitable for mobile devices, which performs tongue segmentation rapidly in an open environment. MethodsFirstly, a new OET-NET is constructed based on U-Net combined with a residual soft connection module and salient image fusion module to train the tongue image from a different device. Secondly, a new loss function is established based on the Focal loss. Finally, the number of network parameters, the inferred time are treated as indicators to evaluate the model. At the same time, MIoU, precision, recall, F1-Score and FLOPs are used to compare with different tongue segmentation frameworks. ResultsAccording to the training results, the proposed OET-NET’s parameter number is 7.75 MB, the required time of tongue segmentation is about 59 ms/piece, the MIoU is 96.98% and the FLOPs is 15.50 MFLOPs. Compared with U-Net, OET-NET increased the total number of parameters by 0.39 MB and the inference time by 1 ms/piece, but achieved the best segmentation effect compared with the reference models. ConclusionsAccording to less time consumption and less space, the precision of segmentation results is higher than that of other segmentation models. OET-NET can quickly and accurately extract tongue bodies from the open environment, and its relatively smaller number of model parameters made it suitable for mobile devices. It is potential for OET-NET to be applied to tongue image segmentation on mobile terminals.