Tongue Image Clearness Judgment and Blur Type Identification

Abstract

In the process of tongue image collection, the image is prone to the following degradation: defocus, Gaussian and motion blur. In this paper, the clearness judgment and blur type identification of tongue images are carried out in order to obtain the accurate blur reasons for further image restoration. For this purpose, this paper first uses the convolution operator (Laplace, Sobel) to obtain the average variance of the tongue image after convolution. Although this method can judge the clearness of the image and severity of blurred images well according to the threshold obtained by using "threshold optimization algorithm", it cannot distinguish different types of blurred tongue images. Therefore, based on the deep learning method, this paper further classifies the clear, defocus, Gaussian and motion blurred tongue images. Firstly, we compressed neural networks from the two dimensions of model depth and width. Compared with the original model, the compressed model with a width factor of 16 and a resolution factor of 4 has 0.09% of the parameters of the original model, 0.62% of the calculation amount(FLOPS) and 83.33% of the test time. Secondly, in order to improve the accuracy and robustness of the model, this paper introduces ensemble learning into blur type identification. The model classification accuracy and the average recall are 96.39%, and the average precision is 96.52%, all of which have different magnitudes of improvement.