针对图像盲去模糊的可微分神经网络架构搜索方法

Differentiable Neural Architecture Search Method for Blind Image Deblurring

为了解决设计图像去模糊神经网络依赖大量手工调参的问题,提出一种面向图像盲去模糊的可微分神经网络架构搜索方法。通过设计U型残差搜索空间,将去模糊网络的搜索过程分为9个搜索单元的搜索过程,降低了搜索的复杂度,并设计出一个基于随机游走和最近邻插值的算法,通过模拟相机运动轨迹的方式生成模糊核,进而生成足够的模糊图像用于训练。实验结果表明,该方法明显减少了人工调参的工作量,在GOPRO和Kohler数据集上搜索得到的网络,峰值信噪比相对于基准网络UNet分别提升3.10dB和1.17dB,并接近UNet的推理速度。

The design of neural networks for image deblurring requires heavy work of manual parameter tuning.To address the problem,a differentiable neural network search method for image deblurring is proposed.By designing a U-shaped residual search space,the search of image deblurring networks is simplified into the search of 9 search units.Besides,this paper presents an algorithm based on random walk and nearest-neighbor interpolation.The algorithm can generate blur kernels by simulating the camera motion trajectory,and thus provide enough blur images for training.The experimental results show that the proposed method significantly reduces the workload of manual tuning.The PSNR of the network obtained by the proposed method is 3.10 dB higher than that of UNet on the GOPRO data set and 1.17 dB on the Kohler dataset,while the inference speed of the network is close to that of UNet.