Abstract: To solve the segmentation problem of the non-uniform illumination and the wide range distribution of grayscale gradation in target and background, considering the information on edge of an image is insensitive to the change of light, quoting of gradient entropy to improve Canny algorithm to detect appropriate edge. Using polynomial curved surface fitting of least square method to get threshold surface to segment targets in the uneven background, so an adaptive threshold surface segmentation algorithm with improved edge detection has been proposed. Various non-uniform distribution of grayscale gradation in background images were verified, the experiment results indicated that the algorithm could get correct edge information, and get relatively good segmentation results to the images with non-uniform distribution of grayscale gradation.