Non-reference image quality evaluation is a research hotspot in recent years. At present, the commonly used evaluation algorithms are extracting features from gray space. In order to increase the reflection of the color channel information on the image quality, the mean subtracted contrast normalized (MSCN) coefficients of each channel in the RGB(Red, Green, Blue), LAB(Luminosity, A, B), and HSV(Hue, Saturation, Value) color spaces are extracted, respectively, and fitted through asymmetric generalized Gaussian distribution model. The statistical features of the fitted MSCN coefficients are trained by gradient boosting regression algorithm to obtain a non-reference image quality evaluation model. The predicted scores of each color channel training model and gray space training model are individually compared with subjective scores. The results show that the monotonicity, subjective and objective consistency, and stability of the non-reference image quality evaluation model under some color channels are improved to some extent compared to the gray space. The model trained with the features extracted under the RGB_B channel has the best performance, Pearson related coefficient increases from 0.63 to 0.70.