Weak supervision super-resolution reconstruction model and method based on close-range and close-range scenes

Abstract

The invention discloses a weak supervision super-resolution reconstruction model and method based on close-range and close-range scenes, and the model comprises: a data set construction module which is used for constructing a near-far scene resolution reconstruction data set for training; a training sample loading module which is used for loading training samples; a first generator network which is used for reconstructing the loaded distant view image of the training sample to obtain a distant view high-resolution image, inputting the obtained distant view high-resolution image into the discriminator and inputting the obtained distant view high-resolution image into the second generator network; a second generator network which is used for reconstructing the input long-range high-resolution image to output a long-range low-resolution image, and the long-range image, the long-range high-resolution image output by the first generator network and the long-range low-resolution image outputby the second generator network form a closed loop; and a discriminator which is used for discriminating the distant-view high-resolution image reconstructed in the first generator network from the close-range image of the training sample, and outputting a discrimination result.

Description

technical field [0001] The invention relates to the technical field of computer image processing, in particular to a weak supervision based on far and near view that uses weakly supervised learning to learn far and near view images of traffic road conditions, and utilizes rich detail information of the near view images to realize super-resolution reconstruction of the distant view images Super-resolution reconstruction method and device. Background technique [0002] In the prior art, when a vehicle captures a long-range traffic image in automatic driving and intelligent monitoring, it is often necessary to perform super-resolution reconstruction on the low-resolution image containing license plate number and road sign information to enrich its detailed information. [0003] At present, the super-resolution reconstruction technology of a single natural image has been developed relatively maturely, and a variety of technologies can be realized, such as SRCNN, FSRCNN, RDN, EDS...

AI Technical Summary

Problems solved by technology

[0003] At present, the super-resolution reconstruction technology of a single natural image has been developed relatively maturely, and a variety of technologies can be realized, such as SRCNN, FSRCNN, RDN, EDSR, SRGAN and other supervised super-resolution reconstruction methods based on deep learning, but These methods cannot use unpaired low-resolution images and high-resolution images for super-resolution reconstruction, resulting in unsatisfactory super-resolution reconstruction of low-resolution images when the true value of low-resolution images cannot be obtained in real scenes

[0004] The existing methods such as CinCGAN for super-resolution reconstruction based on weakly supervised learning can effectively solve the problem of super-resolution reconstruction of unpaired single images, but the resolution of the weakly supervised image domain input to the discriminator is also required to be high-resolution. rate, for example, assuming that the data sets required by the existing weakly supervised super-resolution reconstruction method are unpaired low-resolution images and high-resolution images, in this case, it is necessary to perform Weakly supervised super-resolution reconstruction cannot be realized, and the discriminator of the existing weakly supervised super-resolution reconstruction method can only input the same resolution to generate high-resolution images and weakly supervised high-resolution image domain images

Images