Directional content-preserving features are extracted instead of the global ones. It first detects the initial curvature feature points based on the scale interaction model using Gabor filters with various scales in different directions. Then, the iterative geometric method is proposed to exact the content-preserving feature points. Next, by the radon transform, the two-dimensional distributing is mapped to one-dimensional feature vector. Finally, the one-dimensional feature vectors are quantized to get the hashing. It is verified that the proposed algorithm has a better robustness of standard benchmark (e.g. Stirmark) attacks including JPEG compression (QF=5), geometric distortions of cropping (more than 30%), rotation (more than 10 degree), and common signal processing operations. Moreover, malicious content changing manipulations of image data are also detected. The algorithm has a wide application in identification/search of images in large databases and content-based image authentication, etc.