By taking full advantage of the masking characteristics of the HVS (Human Visual System), this paper first deduces a JND (Just Noticeable Differences) threshold matrix based on blocks, and then presents an image-adaptive semi-fragile watermarking technique operating in DWT (Discrete Wavelet Transform) domain. The proposed technique embeds the watermark by modifying coefficients of the vertical and the horizontal detail subbands of image sub-blocks. The embedded watermark is invisible to human eyes and adaptive to the original image characteristics because of the utilization of HVS characteristics. The experimental results show that the proposed algorithm is effective and can distinguish adaptively some non-malicious image operations such as compression from malicious ones made on the watermarked image. The embedded watermark is sensitive to malicious attacks while robust to non-malicious attacks. Moreover, the scheme has a good accuracy in detecting and locating tampered blocks.