Abstract: This paper proposes a novel semi-fragile watermarking scheme, which is robust against regular manipulations, for image authentication. The semi-fragile watermarking scheme extracts the content feature (watermark) from the original image by adaptively gradient partitioning, and inserts this content feature back into the image by modulating the wavelet coefficients. To enhance the robustness and invisibility of this scheme, the adaptive quantization step is calculated according to the local image characteristics. The integrity authentication and tamper detection are implemented by comparing the extracted watermark and the extracted content feature. Experimental result shows that if there is no change in the obtained image, the watermark will be correctly extracted, and thus will pass through the authentication system. This scheme is tolerant of regular manipulations (such as JPEG2000 compression), but malicious changes of the image will result in breaches of the watermark detection. In addition, this scheme can detect the exact locations---the illegal modified blocks.