Research On Image Super-resolution Reconstruction Algorithm Based On Convolutional Neural Network

Image super-resolution is the process of reconstructing one or more low-resolution images to obtain a high-resolution image and is widely used in applications such as medical imaging,security and surveillance imaging.Because learning-based methods can achieve better performance,they have become a hot spot in the research of super-resolution.Among them,the algorithms using convolutional networks have achieved better performance.But the algoritlhm named image super-resolution using deep convolutional network(SRCNN)has some defects such as less network layers,small receptive field.In this work,three super-resolution algorithms based on convolutional neural networks are proposed to solve those problems.The work of thesis includes the following aspects:1.An image super-resolution reconstruction algorithm based on intermediate supervision convolutional neural network is proposed.Supervision layer loss function and reconstruction loss function are defined to solve the vanishing gradients problem and reduce the computational complexity of model training.The model can reconstruct different degree blurred images well.Experimental results show that the proposed algorithm has better clarity and edge sharpness of reconstructed image.The peak signal-to-noise ratio(PSNR)is 0.41 dB higher than that of the SRCNN.Meanwhile,it takes less time when using the trained network models to reconstruct images.2.An image super-resolution reconstruction algorithm based on residual dense skip connections convolutional neural netvwork is proposed.By using of dense convolution network and residual network connection,it effectively alleviates the problem that information flow is seriously hindered between layers caused by the chain connection in traditional convolutional network.It also realizes the feature reuse,maximizes the information flow and improves the training efficiency.Experimental results show that the proposed algorithm has better clarity and edge sharpness of reconstructed image.The PSNR is 1.04 dB higher than that of the SRCNN.And the shortest running time is 0.03s.3.An image super-resolution algoritlhm based on convolutional neural network with symmetrical nested residual connections is proposed.Its symmetrical combinations provide multiple short paths for the information flow between network layers.The features of different level are fused through skip connections.These symmetrical connections effectively solve the vanishing gradients problem and can implicitly provide guidance for the training of preceding layers,which can be considered as a kind of implicit deep supervision.Experimental results show that the proposed algorithm has better clarity and edge sharpness of reconstructed image.The PSNR is 1.16 dB higher than that of the SRCNN.And the shortest running time is 0.29s.4.By comparing the three proposed algorithms,a modified algorithm for steel cord conveyor belt X-ray images reconstruction is proposed.This algorithm is significantly improved the quality of image reconstruction and can meet the real-time requirements.