Research And Implementation Of Multi-model Super-resolution Framework

With the development of computer vision and deep learning,super-resolution technology is developing rapidly.Super-resolution(SR)reconstruction technology can recover high-resolution images from one or more low-resolution images.SR technology has been successfully used in many fields closely related to image processing,such as surveillance,satellite images,medical images,target recognition and tracking.This paper improves on the model based on generative adversarial networks,and designs a multi-model super-resolution framework that can classify images for training and reconstruction.The multi-model super-resolution framework greatly reduces the model training time and image reconstruction time while maintaining the image reconstruction quality.The trained multiple models can effectively process various types of images.The main work of this article is as follows:(1)We use six different depth models of SRGAN and ESRGAN that have been trained to test 100 images and measure the perceptual quality PI value of the images.We found that among the 6 different depth models of ESRGAN,17 images have the best reconstruction quality in the shallow depth model,27 images have the best image reconstruction quality in the medium depth model,24 image has the best reconstruction quality in the deeper model.Therefore,we conclude that it is not the deeper the model,the better the reconstruction quality.And the shallow depth models can also achieve better image reconstruction quality.(2)In deep learning,the shallow layers of the network is mainly to extract simple features of the image such as texture,color,edge,etc,and the deeper part to extract complex features.Using these features,image recognition and image classification accuracy will be higher.Therefore,we designed an image pre-classifier based on the TVAT standard and using the Sobel operator.The pre-classifier can select the images into three categories: simple,medium,and complex according to the complexity of the image content.The accuracy rate is 71%,and the time overhead is very low.(3)The multi-model super-resolution framework we design uses image classification before training and reconstruction.The image pre-classifier selects the images into three categories: simple,medium,and complex.The simple images are trained and reconstructed in shallow depth model,the medium images category are trained and reconstructed in medium depth model,and the complex images category are trained and reconstructed in high-depth models.Compared with ESRGAN,MMSR can save the average training time by 22.68% and the reconstruction time by 36.41% while maintaining the quality of image reconstruction.