Research On Multi-scale Sub-networks And Combined Prior Information Single Image Super-resolution

Due to the limitations of the size?weight and cost of the imaging devices,the resolution of the captured images is relatively low,which greatly reduces the sharpness of images.At the same time,there is an increasing demand for high-resolution images.Image super-resolution is a typical solution for this problem,whose goal is to reconstruct high-resolution images from low-resolution images.In recent years,abundant image super-resolution algorithms have adopted lots of images to train their deep convolutional neural networks to improved the generalization performance of the model.However,only a few of them focus on the application of applying multi-scale structures to separate and reconstruct the different frequency information.And most algorithms ignore the guidance of prior information for image super-resolution reconstruction.In terms of the above concerns,this thesis studies the application of multi-scale sub-networks model and the models with prior information in image super-resolution,and carries out the following three tasks:(1)By analyzing the current super-resolution algorithms based on convolutional neural network,it is found that a large number of them increase the depth of the network by residual connection and dense connection to improve the reconstruction performance.These methods use a single-scale model to extract features both in high-frequency and low-frequency without separating and how to effectively reconstructing more high-frequency information lacking in the low-resolution image is the key point of this task.Therefore,this thesis proposes a multi-scale subnetworks modle to extract high-frequency and low-frequency information in different sub-networks.In addition,the model uses an adaptive fusion module to improve the fusion of different frequency information.(2)In order to prove the guidance of image prior information on super-resolution reconstruction tasks,this thesis proposes a super-resolution model with image caption.Based on the conditional generative adversarial network with image captions and low-resolution images as in-puts.The model also adds image captions to both the generator and discriminator to overcome the lack of effective semantics of low-resolution images.Meanwhile,the constraint between the reconstructed image and the semantic information is increased by the positive and negative match-aware loss in this model.(3)This thesis uses the facial attribute to guide the reconstruction of low-resolution face images and proposes the image super-resolution method with facial attribute.Different from previous attribute-based face image super-resolution models,this method fuses the attribute information and the facial structure information extracted by the auto-encoder.At the same time,the problem of gradient disappearance is overcomed through the residual and dense connections,hence,deep-structure model can be constructed.The super-resolove images of above three models achieve better evaluation index and vi-sual results against the state-of-the-art methods,and the effectiveness of the model design is demonstrated through model component ablation experiment and analysis experiment.In sum-mary,these three neural-network based methods have improve the reconstruction accuracy of super-resolution and are valuable in research.