基于集成学习的联合去噪去马赛克方法研究

When demosaicing of noisy color filter array(CFA) images, noise pattern form false edge structures, sharpening and amplify high frequency noise, increase the difficulty of denoising. In order to solve the problem of denoising and demosaicing simultaneous, a joint denoising and demosaicing method is designed, based on rough set, support vector machine(SVM) and ensemble learning, which can obtain full color image efficiently. Firstly, build the original sample set for demosaicing, and design attribute reduction algorithm based on rough set. Then, obtain the trained individual supported vector machine using the reduced training samples which are selected from the original sample set. Secondly, design the support vector machine ensemble method, which was used to realize prediction of the missing pixel values, thus, we can get the noisy image. Thirdly, construct original sample set for denoising using multi-feature extractive method by research the description power on the image of different characteristics, then, achieve the classification between the noise point and non-noise point using the proposed support vector machine ensemble method. Finally, establish training samples within the designed neighborhood, then training the γ-support vector regression machine and obtain forecasting model, which can be used to estimate the pixels value of the noise points, and ultimately get the full color image. Therefore, the results of the project research can expand the diversity between the individual SVM, reduce the computational complexity and shorten the training time. Simultaneously, it can get rid of image noises effectively, also can protect edge structure better when interpolation, guarantee obtain the high quality images.

彩色滤波阵列（CFA）图像均为含噪图像，在去马赛克时，噪声模式会形成虚假的边缘结构，锐化且放大高频噪声，增加去噪的难度。本项目针对同时去噪和去马赛克的技术难题，采用粗糙集、支持向量机和集成学习，设计了一种联合去噪去马赛克的方法，能够高效的获得全彩色图像。首先，构建去马赛克的原始样本集，设计基于粗糙集的属性约简算法，获得约简后样本训练出的成员支持向量机；其次，设计支持向量机集成方法，预测待插值点的像素值，得到含噪图像；再次，确定针对含噪图像的综合特征提取方案，构建去噪的原始样本集，以支持向量机集成方法识别含噪图像的噪声点；最后，确定邻域范围并建立训练样本，获得γ-支持向量回归机预测模型，实现噪声点像素值的预测，最终得到全彩色图像。项目的研究将会扩大成员支持向量机之间的差异性，降低计算复杂度并缩短训练时间，在有效的去除图像噪声的同时，还可以在插值时更好的保护图像边缘结构，保证高质量图像的获取。

单传感器的图像采集装置获得的是彩色滤波阵列（CFA）图像，且采集的过程中必然会引入噪声，必须经过去噪、去马赛克的处理才能获得高质量的彩色图像。项目以多技术结合为主线，围绕图像去噪、去马赛克的主题展开研究工作。针对去马赛克时噪声会引起锯齿、伪彩色等问题，利用有限差分、图像相关性和图像梯度等信息设计去马赛克方法，实现了彩色图像的获取；针对SVM的分类和回归两大功能，分别设计了样本构造方案；针对去噪时会引起边缘模糊的问题，设计成员SVM的集成方案实现高精度的噪点检测，再结合分数阶积分算子设计γ-支持向量回归机实现噪点的像素值估计；为了在去噪时更好的保护图像的纹理信息，设计了可变阈值的小波去噪方法，提出了卷积神经网络去噪模型，大大提高了图像的去噪性能。研究成果可以为数码相机、机器视觉中的图像采集装置等的内部软件开发提供理论指导，可以通过以软代硬的方案降低设备成本，具有重要的研究意义。项目发表论文7篇，其中SCI、EI收录6篇，授权发明专利2项，进入实审的发明专利5项，进入公布的发明专利1项，获软件著作权2项，参加国内、国际学术会议7人次，邀请国内外学者作学术报告2人次，培养硕士生4名，晋升副教授2名。