基于颜色空间变换和CNN的自适应去模糊方法

目的研究数字图像中的去模糊问题,从受损的模糊图像中恢复出清晰图像。方法针对现有图像去模糊算法无法保留图像高频信息及容易产生振铃效应等问题,提出一种基于Y通道反卷积和卷积神经网络的两阶段自适应去模糊算法(SDYCNN)。在第1阶段,将数字图像转换至YUV颜色空间,根据图像无参考质量评价分数与模糊核尺寸之间的对应关系,在Y通道内自适应确定模糊核尺寸并进行反卷积增强;第2阶段将第1阶段中的反卷积增强作为预处理方式,通过4层卷积神经网络建立反卷积增强后的图像与清晰图像之间的映射关系,实现图像去模糊。结果轻微模糊图像在第1阶段便能够得到较好的去模糊效果,严重模糊图像经过第1阶段的反卷积增强,也有助于神经网络中特征的快速提取。结论实验结果表明,该算法不仅对于模糊图像具有良好的恢复效果,运算效率也有显著提升。

Self-adaptive Deblurring Algorithm Based on Color Space Conversion and CNN

The work aims to research the deblurring methods in digital image processing that restore clear images from damaged blurred ones. A two-stage self-adaptive deblurring algorithm based on deconvolution and convolutional neural network in Y channel (SDYCNN) was proposed, with respect to the problem that the existing deblurring algorithm could not preserve the high frequency information of the images and was likely to result in ringing effect. In the first stage, the digital images were transferred to YUV color space. According to the correspondence relationship between the image with no reference quality assessment scores and the blur kernel size, the blur kernel size was adaptively determined and the deconvolution enhancement was done in Y channel. In the second stage, the deconvolution enhancement in the first stage was taken as the preprocessing method. The mapping relationship between the deconvolution-enhanced images and ground truth images was established by 4-layer convolutional neural network and could be used for image deblurring. The first stage was sufficient to achieve satisfactory results for slightly blurred image. The severely blurred images after the deconvolution enhancement in the first stage were also helpful for the rapid feature extraction in neural network. Experiment results demonstrate that the proposed algorithm can not only properly restore the blurred images, but also significantly improve the compute efficiency.