As the human eye on the image of different regions of the contrast sensitivity is different, it is particularly important to segment the image region more accurately in the image quality evaluation. Based on this, this paper presents a non-reference image region division method based on deep learning. Firstly, the Canny operator performs image edge detection at low threshold to obtain the strong edge (edge 1) and weak edge (texture) of the image. And then, using edge detection of sparse dictionary (EDOSD) model to get edge 2, which is used to match with the previous edge 1 to choose appropriate segmentation threshold to segment edge and texture more accurately. Finally, the gray image is checked to get flat region of the image. Compared with other algorithms, this method can accurately segment image regions without original image. It is a better strategy to meet the needs of people.