基于Gabor基算法展开卷积神经网络图像恢复模型

Image Restoration Based on Gabor—Based Algorithm Unrolling Convolutional Neural Network

摘要: 基于变分的图像恢复模型可设计高效的即插即用算法，但涉及较多的超参数调节，且去噪器模块通常选择传统深度学习方法，网络层数和训练参数较多。针对此问题，在变分模型算法展开——半二次分裂方法基础上采用径向基函数近似非线性扩散模型的激活函数、采用Gabor基近似卷积核，进而设计了高效紧致的图像恢复深度学习卷积神经网络。Gabor基比离散余弦变换基包含更丰富的方向、尺度信息，可设计高效、紧致的网络结构。与非线性反应扩散模型相比，模型针对图像去噪在噪声级为15、25、50对应的峰值信噪比高出0.14 dB、0.7 dB、0.7 dB，在其它图像问题中也展现出良好的恢复效果。提出的框架不仅能够推广到其它图像恢复任务，还可拓展到其它图像分析与处理的深度学习卷积神经网络设计。

Abstract: The image restoration model based on variational can design an efficient plug and play algorithm, but it involves more hyperparameter adjustment, and the denoising module usually chooses the traditional deep learning method, which has more network layers and training parameters. To solve this problem, an efficient and compact deep learning convolutional neural network for image restoration is designed by using radial basis function to approximate the activation function of non-linear diffusion model and Gabor basis to approximate convolution kernel on the basis of variational model algorithm expansion—semi-quadratic splitting method. Gabor contains more direction and scale information than discrete cosine transform function, which can design efficient and compact network structure. Compared with the nonlinear reaction-diffusion model, the peak signal-to-noise ratio of the model for image denoising is 0.14 dB, 0.7 dB and 0.7 dB higher when the noise level is 15, 25 and 50, which also shows a good recovery effect in other image problems. The proposed framework can be extended not only to other image restoration tasks, but also to deep learning convolutional neural network design for image analysis and processing.