Unsymmetrically trimmed mean filter for noise removal of robot vision in dark environments

Overexposure is likely to occur in the vision system of robots in dark environments, which will result in noise points in the image or video. These noise points caused by overexposure are mostly salt and pepper noise, which may disturb the regular tasks of robots significantly, e.g., target tracking or object recognition. In this paper, a novel unsymmetrically trimmed mean filter (UTMF) is proposed to restore dark-captured images corrupted by salt-and-pepper noise. Firstly, the noise points are detected according to the extreme characteristic of the salt-and-pepper noise. Then, we propose a method to check whether the local pixels around a noise point are smooth or saltant via statistics of difference. Different filtering strategies are designed for smooth and saltant conditions, respectively. Extensive experiments are implemented with standard images to test the proposed method with various noise ratios from 30% to 70%, and the results verify that our method outperforms state-of-the-art methods in both peak-signal-to-noise ratio and structural similarity scores. The peak-signal-to-noise ratio of our method outperforms state-of-the-art methods by 0.1-1.2db. The images captured in dark environments with significant noise points are used to test the proposed method as well, and the results verify that our method is capable of both denoising and detail-preserving.