Face super-resolution system based on multi-scale convolution and receptive field feature fusion

Abstract

The invention discloses a face super-resolution system based on multi-scale convolution and receptive field feature fusion. The system comprises a coarse super-resolution module, a coarse up-samplingmodule, a first multi-scale convolution module, a high and low frequency enhancement module, a fine super-resolution module, a fine up-sampling module, a third-degree scale convolution module, an image enhancement module and an adversarial network which are connected in sequence. A face key point extraction network and the high and low frequency enhancement module are respectively connected with the fusion module, and the fusion module and the coarse super-resolution module are respectively connected with the fine super-resolution module. The system is suitable for face enhancement, especiallyfor a small-resolution face, and has the characteristics of high magnification times and high high-frequency detail reduction degree through high and low frequency feature interaction enhancement andface priori knowledge utilization; and the adoption of the receptive field module is helpful for extracting detail features and reducing the calculation complexity.

Description

technical field [0001] The invention belongs to the fields of computer vision and image processing, and in particular relates to a face super-resolution system based on multi-scale convolution and fusion of receptive field features. Background technique [0002] The size of the image resolution is directly related to the quality of the image. Higher resolution means more detailed information and greater application potential. However, in the actual image acquisition process, due to the limitations of the imaging device itself and the influence of environmental factors, it is impossible to directly acquire high-quality images. At the same time, due to the influence of storage media and network bandwidth, the final image resolution is also low, which hinders Further processing and application of images. With the continuous development of computer vision technology, especially the development of deep learning, there are more and more image quality enhancement methods. Super-re...

AI Technical Summary

Problems solved by technology

[0003] At present, there are many networks for image super-resolution, which have significantly improved in handling various scenes and objects. There are few networks for face super-resolution, and many methods are based on constructing corresponding faces. Data, and then use the existing network for training. Although some progress has been made, the super-resolution effect is not very good for low-resolution faces, and faces have relatively uniform structural information compared to scenes and objects. , low-frequency structural information plays a very important role in face recognition. In order to overcome the shortcomings of existing technologies and improve the effect of face super-resolution, the face is divided into low-frequency structural components and high-frequency detail components, which is different from existing methods. The whole image is integrated into the network, but the input image is decomposed into two parts (structure and details) and sent to the network, and the interactive relationship between these two kinds of information is considered, so that it not only helps to sharpen the face Structural information while helping to recover details

Images