Randomly valued impulse noise removal using Gaussian curvature of image surface

A novel impulse noise removal using Gaussian curvature is proposed. It detects sharp-pointed peaks or steep valleys on the image surface by calculating Gaussian curvatures at all pixels and removes them by using TV inpainting. Generally, the curvature tends to take positive and extremely large values at peaks and valleys. We can take advantage of this property and use it to accurately detect impulse noise by applying several thresholds iteratively in decreasing order. Results of numerical experiments showed that the proposed method is superior to conventional methods in terms of both noise detection and restored image quality, especially in case of randomly valued impulse.