Textile Color Correction Based On Machine Learning

Unstable illumination can result in serious color difference evaluation errors in the dying dyeing process and affect the quality of dyed products. In the traditional textile industry, color correction of dyed products mainly depends on workers with rich experience in color vision. This manual method has many drawbacks, for example, high misjudgment rate, subjectivity, high cost and huge waste of resources. Therefore, it makes sense to estimate the illumination color automatically and efficiently by using machine learning method. The research work of this paper is focus on color correction algorithms under a single light source. We use machine learning technique to construct illumination color estimation models with good generalization ability and stability to solve the problem of color correction of textile in printing and dying process. The main work and achievements of this dissertation are summarized as follows:(1) Firstly, the basic concepts and research status of the color correction are briefly introduced, and then the advantages and disadvantages of various color correction algorithms are compared and analyzed. The paper focuses on the color correction algorithm with better performance, which is based on Extreme Learning Machine(ELM). Finally, the construction and existing problems of ELM are analyzed and discussed, which provides theoretical basis for future work of this paper.(2) In ELM, the input weights are chosen randomly and the output weights are calculated analytically. However, the random selection of learning parameters sometimes produces some un-optimal input weights and hidden biases, which would cause no solution of the linear system. In order to overcome the drawbacks, a hybrid evolutionary approach is presented which uses PSO algorithm to select the optimal input weights and hidden biases for ELM such that the prediction accuracy of base learner ELM is improved. Moreover, in order to further improve the performance of illumination estimation model, Bagging ensemble algorithm is used to generate several different PSO-ELM sub networks. Finally, these individual models are average aggregated in an ensemble model, namely Bagging-PSO-ELM. Experimental results show that the proposed method is more stable and accurate than original ELM and achieves better color constancy than several related methods.(3) In traditional support vector regression(SVR) based color constancy algorithm, the learning speed is very slow due to the single output of SVR and its implementation demands complicated parameter optimization to obtain the best solutions. In order to overcome the drawbacks, we propose a novel illumination correction model based on kernel extreme learning machine(KELM). Furthermore, different from the traditional binarized chromaticity histogram method, we use the Grey-Edge algorithm to extract a low-dimensional and efficient color feature, which is then used as the input vector for KELM. Compared with traditional SVR based color constancy algorithm, the proposed method provides better generalization performance at a much faster learning speed.