To solve the problem of the color difference caused by fabric diversity in the digital printing process， an improved regularization extreme learning machine algorithm is proposed to realize quick and flexible conversion from L*a*b* to CMYK color space. First， select the PANTONE textile TCX color card as the sample data of the experiment， randomly select 800 color patches， the L*a*b* values of these color patches are used as input， and the corresponding C， M， Y， and K values are used as output， respectively. The network is trained to establish a nonlinear mapping， and the optimal regularized penalty coefficient is obtained according to the ridge trace graph observation method of the ridge regression model to optimize the model. Then， 100 color patches are randomly selected from the remaining color patches of the TCX color card as the test samples of the model for testing and verification. Experimental results show that the proposed method has high conversion accuracy and efficiency. The minimum conversion color difference is 0.221， the maximum conversion color difference is 6.965， the average conversion color difference is 1.645， and the average training time is 1.489 s， which can meet the actual requirements of digital printing color management.