An SA-GA-BP neural network-based color correction algorithm for TCM tongue images
Li Zhuo,Jing Zhang,Pei Dong,Yingdi Zhao,Bo Peng
DOI: https://doi.org/10.1016/j.neucom.2012.12.080
IF: 6
2014-01-01
Neurocomputing
Abstract:Tongue inspection is an essential part in the four diagnostic methods in traditional Chinese medicine (TCM). Subject to the variation in conditions such as the imperfection of capturing environment, illumination and imaging devices, the captured tongue images usually contain certain color distortion compared to the actual tongue images, and such distortion has negative impact on the diagnosis from doctors. Therefore, this paper proposes a simulated annealing (SA)-genetic algorithm (GA)-back propagation (BP) neural network-based color correction algorithm for TCM tongue images. The main contributions of this paper include two aspects: First, not all of the color samples from the whole color gamut are used to train the color correction model, only a number of colors that are similar to those of the tongue body, tongue coating and skin are selected from the entire color sample set and used for the color correction, which will greatly reduce the computational complexity of training process and improve the correction accuracy. Second, to further improve the correction accuracy, SA-GA-BP neural network algorithm is utilized in training process to establish the color mapping model, with the captured samples of such color checkers under the capturing environment taken as the input data and the standard color data as output. As to the problem that the color correction models obtained by using the SA-GA-BP neural network method is not unique, the optimal color mapping model is selected based on the principle of minimizing the average color difference between the output values of test samples and standard colors. Experimental results demonstrate that the performance of color correction obtained by the proposed algorithm is superior to that based on the whole color gamut color correction algorithm, while the training time is as 6.7% low as that of the whole color gamut color correction algorithm.