A novel approach of regional blur identification is proposed and applied to image forgery detection. Based on least-square (LS) estimation, prediction error of each pixel is calculated. Variance of the prediction errors of a pixel and its neighborhood is regarded as the blur characteristic of the pixel. A threshold is determined dynamically according to the distribution of correlation coefficient in the frequency domain and blur characteristic in the spatial domain. Thus pixels in a blurred area are discriminated from sharp pixels. Experimental results show that the proposed method has a high accuracy rate in blur region detection.