This paper presents a novel feature point based image watermarking scheme to achieve high capacity information hiding and generalized watermark robustness. The key idea is to embed a binary watermark image into multi-scale feature point based local characteristic regions in transform domain. Watermark synchronization is first achieved by the characteristic regions, which can be extracted using the scale-invariant feature transform and image normalization. Then the watermark image is embedded in a content-based manner by modifying the wavelet transform coefficients. In the detector, the watermark can be extracted from the distorted image directly. Experimental results on Stirmark benchmark 4.0 show that the proposed scheme can resist both traditional signal processing attacks and geometric attacks. Comparisons also demonstrate the advantages of the scheme.