Research And Application Of Differentiable Network Architecture Search In Image Super-resolution

Compared with machine learning methods based on interpolation or reconstruction,deep learning shows its unique advantages in super-resolution(SR)tasks.However,manually designing a super-resolution network with excellent performance often requires a lot of manpower and material resources.In addition,as artificially designed neural networks become more and more complex,it is very difficult to artificially design a well-performing neural network.In order to reduce the manpower and material resources spent by artificially designing neural networks,the value of Neural Architecture Search(NAS)has emerged.Neural architecture search methods based on reinforcement learning,evolutionary algorithms,and Bayesian optimization strategies are gradually replaced by methods based on gradient descent algorithms because they consume huge GPU resources.Differentiable Neural Architecture Search(DARTS)not only does not include the controller,super-network and model predictor,but also reduces the network architecture search time to a few GPU days.NAS is rarely used in super-resolution task.In order to automatically design a well-performing super-resolution network architecture,differentiable neural architecture search searches for super-resolution architecture.The main work is listed as follows:(1)Reconstruct the search space of the differentiable neural architecture searchDARTS was originally used for image classification and image recognition tasks,so the search space of DARTS is not necessarily suitable for single image super-resolution task.After studying the network architectures of the super-resolution classic model EDSR,MDSR,etc.,it is found that these networks are all obtained by stacking Resblock or a variant of this module.From this analysis,it can be concluded that the super-resolution task requires not only the high-dimensional information extracted by the convolution operation,but also the low-dimensional information provided by the identity mapping.Therefore,this paper adds an identity mapping operation to the convolution operation of the neural network architecture search space to retain the low-dimensional information of the feature map.(2)Application of improved differentiable neural architecture search in super-resolutionExperiments have shown that the direct use of DARTS on super-resolution tasks will cause too many skip-connect operations in the searched super-resolution network architecture,resulting in poor super-resolution effects in the final network architecture.Based on the reconstructed search space of the differentiable neural architecture,this paper uses the differentiable neural architecture search to search for the cell structure,and stacks the obtained cells to form a super network.The network is used as the non-linear mapping part of the super-resolution network to extract high-dimensional semantic information.Experimental verification shows that the super-resolution network architecture searched by the above method is effective on the benchmark data set and DIV2K data set.(3)Low-latency super-resolution network architecture searchOn the basis of the Three Freedom NAS(TF-NAS)search,this paper designs a new method to automatically search out the delay-constrained super-resolution network.Because the inverted residual block proposed in MobileNetV2 can reuse image features and improve network performance,this operation is used as a basic candidate operation in the neural architecture search space,which can greatly reduce the number of network parameters and the running time of the network architecture.At the same time,in order to ensure the low latency of network inference,the dimension of the middle layer is dynamically adjusted to make the latency of the network architecture close to the target latency.Quantitative and qualitative evaluations on the benchmark datasets show that the network latency obtained by the proposed method is low and the generated image quality is high.