A Study of Noise Detection and Restoration Techniques for Corrupted Images

• Main Authors: Hwang-Yu Chen, 陳煌玉
• Other Authors: Hsien-Chu Wu
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/2a98ta

碩士 === 臺中技術學院 === 資訊科技與應用研究所 === 97 === In this thesis, three schemes of noise detection and restoration techniques are proposed. The proposed schemes mainly analyze spatial characteristics of noises for digital images to enhance the accuracy of traditional noise detection schemes. In order to obtain a restoration image which is close to the original image as possible, a proper filter can be employed to recover the corrupted image. The first scheme for random-value impulse noise removal is proposed by employing the iterations, variances, predefined threshold values and weight techniques to progressively improve the quality of noise detection and restoration image. The scheme counts the number of the pixel value differences which are greater than the predefined threshold, and these differences between the center pixel and its eight neighbor pixels in the sliding window. When the count is over half of the window size, the evaluated pixel is classified as noise one. The modified directional weighted median filter is applied to restore the corrupted image against detected pixels at first. This scheme progressively exploits noise detection and restoration in order to obtain the higher quality and less distortion for restoration images. In several experiments on various kinds of images, the proposed method acquires a group of best thresholds that is proved to be suitable for numerous types of images. The experimental results show that the proposed scheme can effectively detect noises, preserve edge details and improve the quality of restoration images. The second proposed scheme mainly exploits the boundary technique to detect noise pixels and utilizes the interpolating technique to recover the corrupted images with the impulse noises. The boundary technique is modified to obtain more efficiency and accuracy of noise detection. It can combine with the proposed interpolating technique in the sliding window to extract noise-free pixels from four defined directions of probable edges in order, and then these found pixels are employed to estimate the original value for each detected pixel. The proposed scheme can achieve both quick processing and higher quality of restoration images. Similarly, the scheme also exploits the gradual evolution way to reduce the distortion and enhance the restored quality of the corrupted images. From the experimental results, the proposed scheme has highly accurate noise detection and achieves zero miss-detection rate across a wide rage of noise densities, from 10%-90%; and then it can be incorporated with the interpolating technique to achieve an estimation value close to the original pixel-value and enhance the quality of the entire image. The third scheme is proposed for impulse noise removal similarly by the basic morphological operations to quickly judge whether pixels are noises and using a direction weight average value technology to restore corrupted images. This study modifies the judged condition of morphology to improve noise detection rate. Normal pixels can be found from twelve defined directions of probable efficient edge in order, and one pair of pixels lay in the most possible edge is weighted in the filtering window, and then obtains the estimation value by averaging those extraction values. The scheme achieves a less distortion and low computation time by using a little sliding window in the restored processes. The proposed scheme takes the single iteration as the processing principle, but it also restore image by the gradual way to enhance the restoration quality under the high interference, especially. According to the experimental results, the proposed scheme takes both quick and efficient to detect impulse noises that achieve zero miss-detection rate and nearly are close to zero false-detection across a wide range of noise densities, ranging from 10%-90%; and then combine the proposed edge directions with weighted mean filter to achieve a highly restoration by using only a little filtering window.