Unsupervised Real Image Super-resolution Via Knowledge Distillation Network

High resolution image can provide abundant spatial structure information,which is an important development direction of color and spectral imaging equipment.However,affected by the hardware limitations of imaging equipment,noise,and compression during transmission,the obtained image resolution is usually not ideal,which easy to cause some key information of the target in the imaged scene to be lost.Therefore,under the above imaging conditions,it is of great meaning for information mining and utilization to improve the resolution of the image,namely,the image super-resolution reconstruction technology.Over the last years,image super-resolution has become an important research aspect in computer vision due to its important applications in monitoring and recognition,unmanned driving,and medical imaging.Super-resolution aims to convert a given low-resolution image with coarse details to a corresponding high-resolution image with higher visual quality and refined details.At present,with the rapid development of deep learning,image super-resolution based on deep learning has also made many research results in theory and application,but there are few data,paired images are difficult to obtain,the true degradation process and the reconstruction process degradation kernel mismatch and other issues.but there are some problems such as less data,difficult acquisition of paired images,mismatch of degraded kernels in the degradation process and the real degradation process.Therefore,how to better restore high-resolution images in real scenes is a challenging and open topic.In order to better solve the task of image super-resolution reconstruction in real scenes,this paper only uses a single low-resolution test image to study from the perspective of unsupervised image super-resolution,introduces a knowledge distillation network,and proposes an unsupervised super-resolution network model based on knowledge distillation,which achieves high super-resolution reconstruction performance of real scene images.The primary work of this paper is as follows:(1)Zero-shot image SR model deals with different images based on the same degradation model,which is inconsistent with the real degradation process.To solve the above problems,an adaptive image degradation model is proposed by using the self-similarity of images.By mining the block similarity of low-resolution test image and its degraded image,the specific degraded kernel and corresponding degraded image are obtained.(2)The mapping between the low-resolution degraded image and the low-resolution image is not completely equivalent to the mapping between the low-resolution image and the high-resolution image.Therefore,this paper proposes the use of knowledge distillation network to transfer the mapping learned from the teacher network to the process of reconstructing high-resolution images from low-resolution images,which can effectively improve the image detail recovery ability of the super-resolution network.In addition,when the low-resolution degraded image generated by the degradation model is used as the input of the teacher network,the available information in the image will be reduced,resulting in the learned mapping is not accurate enough.To solve the above problems,a preprocessing module based on residual learning is proposed to assist knowledge distillation.(3)In this paper,the image super-resolution network model established is tested on synthetic images and real images,and good performance of image super-resolution reconstruction is obtained.The experimental results show that the algorithm can reconstruct the details of the image well in the real scene,which provides an effective method to generate informative learning tasks with fewer samples.