DeepWPD: A deep learning based wavelet packet decomposition to remove Moiré pattern from screen capture images.

Abstract

Moiré artifacts is a special type of noise which is rarely considered in deep learning based image processing tasks. But with the increasing number of digital screens like TV, laptop, desktop screens etc. it is becoming common to take pictures of these screens to quickly save information and a common aliasing effect in these screen cap ture images is moiré pattern. These kinds of artifacts in images appear when two repetitive patterns interfere with one another. Moiré patterns degrade the quality of photos. It affects the performance of other deep learning tasks using these images like classification, segmentation etc. As the moiré pattern is highly variant, has im balanced magnitude in different channels and sophisticated frequency distribution, they are difficult to be completely removed without affecting the main information of the underneath image. Because of its complex nature, most state-of-the-art im age restoration and denoising related methods fail to remove these artifacts. In this paper I proposed an effective wavelet based deep learning model for removing moiré patterns which outperforms all other state of the art by large margins. My pro posed model recovers the details of the moiré free image using the wavelet packet transform. The Residual Dense Module Network and Dilation Convolution Network of our model acquires moiré information from almost all frequency ranges. One for high frequency range and other for low frequency range.