Demosaicking is the way toward reproducing a full hued picture from a deficient shaded picture. The single sensor doesn't catch all hues for a single pixel. To address this, a color filter array (CFA) is utilized to get a hued picture from a single sensor. The created picture from CFA is called a mosaic picture. In this research, we utilize specialized networks to remove the noise from Bayer images. The mosaic picture is adulterated by commotion presented by a sensor or other equipment during catching. Demosaicking on the boisterous mosaic picture makes antiquities, for example, moire and zippering. Some solutions have been proposed for denoising mosaic images but they are handcrafted solutions. In this paper, a solution is proposed to the first denoise and then demosaic the image using machine learning. The mosaic image is denoised using CNN which is then demosaicked using the residual learning strategy of a single specialized network. One of the networks is DHTN (deep high textured network) which is trained on textured images and the second one is DSTN (deep smooth textured network) which is trained on smooth images. Preliminary results show that the proposed approach generates better results and higher quality images than traditional approaches.

中文翻译：

---

使用深度学习基于专用网络的拜耳图像去马赛克和去噪

去马赛克是从有缺陷的阴影图片中再现全色调图片的方法。单个传感器无法捕捉单个像素的所有色调。为了解决这个问题，我们使用了彩色滤光片阵列 (CFA) 从单个传感器获取彩色图片。从 CFA 创建的图片称为马赛克图片。在这项研究中，我们利用专门的网络来去除拜耳图像中的噪声。马赛克图片是由传感器或其他设备在捕捉过程中呈现的骚动造成的。在喧闹的马赛克图片上进行去马赛克制作古董，例如莫尔条纹和拉链。已经提出了一些用于马赛克图像去噪的解决方案，但它们是手工制作的解决方案。在本文中，提出了一种使用机器学习对图像进行先去噪然后去马赛克的解决方案。马赛克图像使用 CNN 去噪，然后使用单个专门网络的残差学习策略去马赛克。其中一个网络是在纹理图像上训练的 DHTN（深度高纹理网络），第二个是在平滑图像上训练的 DSTN（深度平滑纹理网络）。初步结果表明，所提出的方法比传统方法产生更好的结果和更高质量的图像。