Neural component searching method for generating image super-resolution network

Abstract

The invention discloses a neural component searching method for generating an image super-resolution network, and the method can search the image super-resolution network under different computing power limits, and the size of the network is scalable. According to the method, besides a neural network structure, color space, loss function composition and other neural components influencing neural network training and reasoning performance can be searched. According to the method, a weight sharing strategy is used at the same time, hypergraph weights constructed in all search spaces can be inherited by the searched sub-graphs, and the search time is greatly shortened. Aiming at image super-resolution reconstruction application, a simple hypergraph is constructed, and slow, general and fast image super-resolution networks are searched. In addition, the method is not only suitable for the field of image super-resolution networks, but also can be used as a general framework, and neural networks of other applications can be searched by using the proposed extension component search method by changing the search space.

Description

technical field [0001] The invention belongs to the technical field of image super-resolution, and in particular relates to a neural component search method for generating an image super-resolution network. Background technique [0002] The current image super-resolution reconstruction methods can be divided into interpolation-based methods and learning-based methods. Interpolation-based methods can use parametric methods, such as bicubic interpolation; non-parametric regression methods such as edge-directed interpolation, normalized convolution, bilateral filtering, etc. can be used to upsample the scale of the image. However, interpolation-based image super-resolution reconstruction methods perform better in smooth regions (low frequencies) and poorer in edge regions (high frequencies), because they are prone to edge blurring and jagged artifacts. [0003] Learning-based image super-resolution reconstruction methods have some good reconstruction results, but most learning...

AI Technical Summary

Problems solved by technology

The problem with manually designed networks based on experience and multiple experimental adjustments is that different networks need to be trained in different super-resolution scenarios, and the networks are different in many aspects such as structure and loss function.

However, in these neural structure search methods, the search space is limited to the network structure, and a large amount of artificial prior knowledge is still introduced in the loss function and training strategy; and most search methods use genetic algorithms, and the search space is large, resulting in long search time

Images