马赛克图像是一类特殊的图像，它通过对清晰图片进行分区块的处理来生成确定的低清图片。对于一张已有的清晰图片，计算机上有许多图像处理软件能提供将其马赛克化的功能。但与此同时，由于马赛克化的处理对图像包含的信息是有所丢失的，因此从一张已经被打上马赛克的图片还原出一张原本的清晰图片是十分复杂的。同时，受制于一直以来计算机计算能力的有限，相关算法的开发也一直受到了硬件条件的制约。

    自本世纪 10 年代以来，随着计算机处理技术的进步，特别是显卡开发能力的提升，一般的计算机在今天也能拥有强大的处理能力，这使得利用深层神经网络模型逼近图像之间的复杂映射成为了可能。从 2012 年深层神经网络模型首次展现出在完成图像处理任务时的巨大优势以来，已有众多研究学者利用深层神经网络模型完成了物体目标探测，图像语意分割等众多复杂且实际的问题。

    基于以上的事实，深度神经网络系统在寻找由马赛克图像还原清晰图像的映射方面也应当有及其显著的优势。事实上，已有相关的研究学者在这一方面发表了优秀且可靠的算法，但受制于马赛克处理形式对图像信息的丢失过多，这样的算法现在仍旧存在一些缺陷。

    本文将在前人研究的基础上，深入探讨有监督学习下的马赛克图像的还原问题。为了实现这一目标，本文参考了在相关领域的研究工作，自主设计了一深层的神经网络模型并选用了合适的损失作为参数迭代的参考。

    在对算法进行设计以及调整后，本文还进行了大量的算法实验工作，并且将实验结果与传统的图像还原方式进行比较。与传统方式相比，本算法的还原效果有显著提升，展现了深层神经网络在相关任务中的巨大优势，并且算法相关的程序能在一般计算机上进行。

    相关的实验结果以及测试数据均在文中给出，根据实验结果和测试数据可以认为该算法比较好的完成了马赛克图像还原的任务。同时，希望这一研究成果能为未来算法的开发带来启发。

Mosaic is a special way to process images.It divides a clear picture into many small pieces and applies a certain function in all pixel points of each pieces to create a blurry image. Today, many computer softwares are able to create a mosaic picture from a given high-resolution picture. Meanwhile, demosaic still remains a complex process since mosaic pictures lack information containing in original pictures. Moreover, due to the shortage of calculation ability, development in related algorithms had been stagnant.

Since 2010s, with the improvement of computer hardware, especially GPU, even an ordinary computer today obtains calculation ability better than any other times. This means developing a deep neural network model to approach complex mappings between different pictures is possible now. Ever since deep neural network models showed its huge advantages in processing images, many scholars have donated themselves in the research of related fields and completed many hard tasks such as object detection and pattern recognition.

Based on the mentioned factors, deep neural network models should be outstanding in searching for demosaic functions. As a matter of fact, some excellent algorithms have been published by scholars in related fields. However, these algorithms still have a few of shortcomings because mosaic loses much information of high-resolution pictures.

Build on the ideas of previous work, this thesis intends to discuss problems about demosaic in supervised dataset. To achieve the goal, influencing by research work in related feilds, a new neural design is proposed and suitable loss functions are chosen.

After designing and adjusting this algoriehm, multiple experiments are performed to test the effciency. Compared to traditional method, the results suggest that neural network model can improve the accuracy of image recovery, which indicates deep neural network models have many advantages in related tasks. The programs of this algorithm can be run in a regular computer.

The results and test data is given in this article. Based on the results and data, this algorithm is considered to be capable for demosaic and its related tasks. Meanwhile, this work is hoped to bring inspiration to future algorithm developments.