Research And Recognition Applications Of Image Super-resolution Algorithms For Real Scene

Image super-resolution algorithms can improve image resolution,recover high-frequency details of images and enhance visual sensory effects under the performance limitation of imaging devices,which is an important branch in the field of computer vision.Current research on image super-resolution algorithms based on deep learning tends to deepen and widen the network to improve performance,but it may also bring problems such as gradient dispersion and feature redundancy,which cause network degradation,and most of the research ignores the full utilization of different scale features,which in turn affects the network reconstruction effect.In addition,the current research mainly uses the ideal degradation model(e.g.,double cubic interpolation downsampling)to obtain low-resolution images,while the image degradation in real scenes is affected by many factors such as noise,blur and mechanical vibration,and the process is complicated and difficult to quantify,which does not conform to the assumption of the ideal degradation model,resulting in serious performance degradation when the superresolution algorithm is applied to real scenes.In this paper,we conduct an in-depth study to address the above issues,and the main work is as follows.(1)We propose an image super-resolution reconstruction algorithm based on multi-scale convolution and residual-dense attention structure.By introducing the multi-scale convolution and residual dense concatenation methods and attention mechanism to improve the feature extraction structure in SRRes Net,a multi-scale residual dense concatenation attention structure is designed to make full use of the multi-scale feature information,promote the screening and circulation of information,and solve the problem of long-term network dependence by fusing local and global features.In the comparison experiments the algorithm proposed in this paper achieves better results in both objective and subjective indicators.(2)Improved image degradation method for fuzzy kernel estimation.The improved downsampling model extracts fuzzy kernel information and noise information from lowresolution images of real scenes,and applies these a priori information to high-resolution images to generate low-resolution images close to real scenes,so as to construct a dataset that matches the distribution of real scenes.It is experimentally demonstrated that the superdivision network trained with the improved degradation method outperforms the network trained with the ideal degradation method in both simulated and real image experiments,effectively improving the application of the super-resolution network in real scenes.(3)The above super-resolution method is applied to the traffic sign recognition scenario.The image quality of traffic sign images is generally not high due to the limitation of real scenes.To address this problem,this paper uses super-resolution techniques to improve the effect of traffic sign recognition and proposes super-graded association networks for targeted optimization of the target task.It is proved through practical applications that super-resolution techniques can not only improve the image quality of real scenes,but also help to improve the effect of other computer vision tasks.