The work aims to propose a robust image watermarking algorithm based on elliptical feature region and important bit-plane decomposition, for the purpose of solving the defects such as low robustness of the local feature region induced by difficultly in describing the non-texture region of the image (the edge with greater pixel intensity changes, uniform area where pixel intensity is smaller or tends to be zero) in current watermarking scheme based on feature points, thus resulting in its poor anti-geometric distortion ability. Firstly, the color invariance was calculated according to the RGB component of the color carrier. The probability density gradient estimation function was derived based on the intensity probability density of color invariance. Subsequently, the probability density gradient was used to compute the two-order Hessian matrix for improving the SURF method, which fully extracted the robust feature points of the texture and non-texture regions in the color image. The eigenvalues and eigenvectors of the Hessian matrix were obtained to calculate the radius of the long axis and short axis with the ellipse, as well as the angle of the direction; and the local elliptical feature region was established by taking the feature points as the center. Then, a robust feature region selection rule was defined to determine the appropriate watermark embedding location. The bit-plane decomposition technique was introduced to get the important bit-plane image of the robust elliptical feature region, and its histogram was extracted. Based on that, the watermark embedding method was designed to hide the binary watermark in these histograms for forming watermark image. Finally, the binary watermark was restored based on watermark detection mechanism. The experimental results showed that the proposed algorithm had higher uncertainty and robustness, as well as the least distortion of the restored watermark, compared with the current watermarking scheme based on feature points. The proposed algorithm has higher visual stealth and anti-geometric distortion ability, which has better reference value in copyright protection, information security and other fields.