Research On Image Super-resolution Reconstruction Technology Based On Deep Learning

Image super-resolution reconstruction technology is becoming a research hot-spot in the field of computer vision.Image super-resolution reconstruction algorithms based on convolutional neural networks have received widespread attention recently.However,many current models and algorithms do not achieve satisfactory results.This thesis improves the algorithms and implements the restoration of high-resolution image.This thesis designs a super-resolution reconstruction algorithm based on deep residual learning.The feature extraction phase consists of cascaded special convolution to expand the receptive field.The feature learning phase utilizes a combination of global residual learning and local residual learning structure.On one hand it is able to mine more internal relationships of image features,on the other hand it can retard the network degradation,gradient disappearance and explosion.The image reconstruction phase applies sub-pixel convolution to reduce external noise interference and meet multi-scale amplification requirements.In addition,this thesis proposes a super-resolution reconstruction algorithm based on the fusion attention mechanism.The algorithm introduces the fused channel attention and spatial attention mechanisms into the residual module,which increase the model's ability to discriminate different features and pay attention to important feature channels and regions.In terms of model structure and parameter training,This thesis utilizes PRe LU instead of Re LU activation function to avoid the problem which negative parameter area is not activated.This thesis adopts MSRA instead of Xavier parameter initialization method to make the parameter distribution of training process more reasonable.This thesis employs WN instead of BN normalization operation to get rid of dependence on training sample data distribution.Moreover,the loss function combined with L1 and L2 can not only accelerates the convergence speed of parameter learning,but also makes the parameters converge to the global optimal solution.Finally,We perform training and testing on standard public data sets.Through various comparative experiments,it is fully verified that the algorithm proposed in this research can reconstruct higher-quality super-resolution images from the perspective of subjective and objective evaluation.