一种基于双分支改良编解码器的图像去噪算法

针对传统图像去噪算法多噪声去除难，深层卷积神经网络去噪模型网络复杂、训练时间长等问题，提出一种基于自编码器结构的双分支改良编解码网络，实现高效图像去噪。双分支结构之一采用降-升采样实现点噪声消除，另一分支专注于宏观的图像修复和伪像去除，后端利用残差结构进行整合，实现数字图像混合噪声去噪。实验结果显示:对于含有标准差为15，均值为0的高斯噪声、噪声密度为5%的椒盐噪声和散粒噪声的混合噪声图像测试集，实验去噪效果相较于输入混合噪声图像峰值信噪比，平均提升了5.3%。与12层全卷积神经网络相比，去噪效果相当，训练速度提升了25.4%，体现了其“轻量级”的优点。实验表明:该方法相较于深层卷积神经网络，训练速度快，网络简单；相较于传统图像去噪算法，噪声去除效果也较为明显。该算法可应用于轻量级视觉平台后端去噪。

Image denoising algorithm based on dual-branch modified codec

Aiming at the problems of the traditional image denoising algorithm such as difficult multi-noise removal, complex deep convolutional neural network denoising model network and long training time, a dual-branch modified codec(DMC) network based on auto-encoder structure was proposed to achieve the high-efficient image denoising. One of the dual branch structure used the down-up sampling to eliminate the point noise, the other focused on the macroscopical image restoration and artifacts removal, and the residual structure was used to integrate at the end to realize the mixed noise denoising of the digital image. The experimental results show that for the image test set of the mixed noise containing Gaussian noise with standard deviation of 15 and mean value of 0, salt and pepper noise as well as shot noise with noise density of 5%, compared with the peak signal-to-noise ratio of the input mixed noise image, the experimental denoising effect is improved by 5.3% on average. Compared with the 12-layer full convolutional neural network, the denoising effect is equivalent and the training speed is increased by about 25.4%, which embodies the advantages of its lightweight. The experimental conclusions indicate that compared with the deep convolution neural network, this method has the advantages of fast training speed and simple network; compared with the traditional image denoising algorithm, it has better noise removal effect. This algorithm can be applied to the end denoising of lightweight vision platform.