基于深度双边格的实时图像增强算法的研究

基于深度双边格的实时图像增强算法的研究(任务书,开题报告,外文翻译,论文19000字)
摘要
性能是移动图像处理算法的重要因素和指标。如今手机或者其他移动设备生成的图像和视频的分辨率越来越高,这给图像处理算法带来了巨大的处理压力,仅专业程序员通过代码优化才勉强解决大分辨率图像处理的效率问题。尽管计算机系统的设计试图高效的运行复杂图像处理算法,但它们需要编程人员的专业知识,其运行时成本仍然随着流水线复杂性的增长而增长。此外,因为图像增强的质量评价具有一定的主观性,在图像增强调整过程中学习颜色改变规律模型。
毕业设计讨论了增强图像的方法。引入了双边网格处理和局部仿射变换相结合的新型神经网络架构。利用输入/输出图像对,训练卷积神经网络来预测双边网格节点的局部仿射模型的系数。在运行时,神经网络模型需要输入低分辨率图像,在双边空间中计算出一组仿射变换,然后把这些经过线性插值返回至全分辨率图像。
所讨论的算法可以生成质量与高质量的图像增强效果图片,而且可以在嵌入式设备上实时运行。所讨论算法传递Google Pixel手机上的全屏分辨率1920ⅹ1080图像仅需要14毫秒,从而实现50帧/秒的实时取景器效果。
关键词:深度学习双边格实时图像增强 [资料来源:https://www.doc163.com]
Research and implementation of deep bilateral learning for real-time image enhancement
Abstract
Performance is an important factor and indicator for image processing algorithms for mobile devices. Nowadays, the resolution of images and video generated by mobile phones or other mobile devices is getting higher and higher, which brings huge processing pressure to image processing algorithms. Only professional programmers can barely solve the problem of high-resolution image processing efficiency through code optimization. Although computer systems are designed to efficiently run complex image processing algorithms, they require the expertise of programmers, and their runtime costs continue to grow with the complexity of the pipeline. In addition, because the quality evaluation of image enhancement has a certain degree of subjectivity, a color change law model is learned during image enhancement adjustment.
Graduation design discusses ways to enhance image. A new neural network architecture combining bilateral grid processing and local affine transformation was introduced. Using input/output image pairs, a convolutional neural network is trained to predict the coefficients of the local affine model of bilateral mesh nodes. At runtime, the neural network model needs to input low-resolution images, compute a set of affine transformations in bilateral space, and then return these to a full-resolution image after linear interpolation. [资料来源:Doc163.com]
The discussed algorithm can generate quality and high quality image enhancement effect pictures, and it can run on embedded devices in real time. The algorithm in question delivers a full-screen resolution of 1920x1080 on a Google Pixel phone in only 14 milliseconds, achieving a real-time viewfinder effect of 50 frames per second.
Keywords: deep learning; bilateral grid; real-time image enhancement
[来源:http://Doc163.com]


目 录
摘 要 1
Abstract 2
第一章 绪论 1
1.1基于深度双边格的实时图像增强算法的研究背景 1
基于GAN的方法1.2基于深度双边格的实时图像增强算法的研究意义 3
1.3目前的相关研究工作 4
1.4双边格的发展方向概况 5
1.5基于深度双边格的实时图像增强算法的相关研究 7
1.6本毕业设计论文的结构 8
第二章 相关基础知识 9 [版权所有:http://DOC163.com]
2.1机器学习开源软件库TensorFlow 9
2.2 TensorFlow在Ubuntu系统中环境的搭建 9
第三章 需求分析 16
3.1低分辨率预测双边系数 17
3.1.1 低级特征 17
3.1.2 局部特征路径 18
3.1.3 全局特征路径 18
3.1.4 融合和线性预测 19
3.2图像特征作为双边网格 19
3.3使用可训练的切片图层进行上采样 20
3.4全分辨率图像插值生成 22
3.4.1 导向图辅助网络 22
3.4.2 组合产生最后结果 24
3.5训练 25
第四章 系统详细设计 26
4.1 Tensoflow核心函数设计 26
4.2 各个功能模块的代码设计 31
4.3 代码运行顺序 34
第五章 结果和性能评估 36
5.1图像操作处理 36
5.2性能 38 [资料来源:https://www.doc163.com]
5.3讨论和局限 40
参考文献 I
致谢 I [来源:http://Doc163.com]
