Research On Image Super Resolution Reconstruction Technology Based On Deep Learning

For various image processing tasks,the quality of the image to be processed directly determines the difficulty and completion of the image processing task.However,due to the environmental conditions of image acquisition and the cost of hardware devices,the resolution of most images does not meet the requirements of subsequent image processing tasks.So it is particularly important to improve the resolution of existing images by software methods.The difficulty is that the low-resolution image predicting the high-resolution image needs to solve an underdetermined equation,which requires a process of complex feature extraction and transformation.The deep-learning method can learn complex feature extraction and transformation methods from big data.It is very suitable for image super-resolution reconstruction tasks.Experiment results prove that the deep learning algorithm has achieved great success in image super-resolution reconstruction.This paper focuses on the shortcomings of the deep learning algorithm in the field of super-resolution in the current stage,and has achieved a series of innovations:Firstly,aiming at the over-fitting problem prevalent in super-resolution convolutional networks,a regularization method based on enhanced prediction is designed to improve the accuracy of SRCNN algorithm.It is confirmed that the super-resolution convolutional neural network does have over-fitting problem.Secondly,the feature extraction layer of the network in FSRCNN algorithm is improved,the depth of the algorithm is improved,and more features can be obtained for nonlinear transformation.Experimental results prove that the feature extraction method of super-resolution convolutional network is indeed worth improving.Thirdly,based on the dense connection structure,a dense and compact network is designed.Under the premise of not losing precision,the operation speed of image super-resolution is greatly improved,the parameter quantity of the model is greatly reduced,and the practical utility of super-resolution convolutional neural network is improved.Finally,a dual path network is designed based on dense connection and residual structure,which deepens the depth of the network and performs more complex feature transformation on low-resolution images,thus improving the accuracy of super-resolution reconstruction.Aiming at the problems existing in the deep resolution reconstruction algorithm based on deep learning,this thesis conducts in-depth theoretical and applied research,and proposes improved method for the accuracy and speed improvement of super-resolution reconstruction.It has contributed to the research of image super-resolution reconstruction algorithm.