基于生成对抗网络的CFA图像去马赛克算法

CFA Image Demosaicing Algorithm Based on Generative Adversarial Network

在单传感器数码相机图像采集系统的彩色滤波阵列中,每个像素仅捕获单一颜色分量,并且在彩色图像重构过程中图像边缘等高频区域的伪影现象尤为明显。提出一种基于生成对抗网络的图像去马赛克算法,通过设计生成对抗网络的生成器、鉴别器和网络损失函数增强学习图像高频信息的能力,其中使用的生成器为具有残差稠密块和远程跳跃连接的深层残差稠密网络,鉴别器由一系列堆叠的卷积单元构成,并且结合对抗性损失、像素损失以及特征感知损失改进网络损失函数,提升网络整体性能。数值实验结果表明,与传统去马赛克算法相比,该算法能更有效减少图像边缘的伪影现象并恢复图像高频信息,生成逼真的重建图像。

The image acquisition system of the single-sensor digital cameras includes a color filtering array which captures only one color component in each pixel,and the artifacts in high-frequency areas such as image edges are particularly noticeable in the reconstruction of colorful images.A CFA image demosaicing method based on Generative Adversarial Network(GAN)is proposed.For the network,a generator,a discriminator and the network loss function are designed to enhance its ability to learn the high-frequency information of images.The generator employs a deep residual dense network which includes residual dense blocks and remote skipping connections.The discriminator consists of a series of stacked convolutional units.In addition,the network loss function is optimized by combining the adversarial loss,the pixel loss and the feature perception loss.The results of the numerical experiments show that compared with the traditional demosaicing algorithms,the proposed algorithm can more effectively reduce the artifacts at the image edges and recover the high-frequency information,providing more vivid reconstructed images.