In order to solve the problem as low location and detection accuracy of duplicate content areas induced by ignoring the color information and the correlation between different color components in current image forgery detection method when identifying the duplicate content area, this paper designs an image forgery detection method based on improved Speeded Up Robust Features(SURF) descriptor and multi-polarity complex exponential transformation. The Gaussian low pass filter is introduced to filter the color image for eliminating noise. Then the color invariance of the filtered image is calculated, and the gray component of the SURF descriptor is replaced by the color invariance to improve the SURF method for obtaining a new Hessian matrix and detecting the interest points in the color image adequately. Subsequently, a set of connected Delaney triangular networks is constructed by using these interest points. The local visual features of triangular networks are extracted based on the quadripolar complex exponential transformation. The Euclidean distance between visual features is calculated, and the triangulation mesh is registered according to the preset threshold. Finally, random sample consistency is introduced to eliminate mismatched triangular networks, and post-processing method is defined to locate the duplicated and forged regions. The test data show that compared with the existing copy-paste forgery detection methods, the proposed method has higher accuracy in forgery detection under various geometric transformations.