Image reconstruction method based on iterative convolutional neural network

Abstract

The invention discloses an image reconstruction method based on an iterative convolutional neural network. The method comprises a multi-scale progressive convolutional neural network model training part and a multi-scale progressive convolutional neural network model testing part; the method includes: in a model training phase, segmenting the images in the public database, calculating the angles of the image blocks, clustering the image blocks according to the angles, constructing high-resolution image clusters and low-resolution image clusters at different angles, training a multi-scale progressive convolutional neural network for the high-resolution image clusters and the low-resolution image clusters at different angles, and solving convolutional neural network model parameters at different angles; in a test phase, segmenting the test low-resolution image into image blocks and calculating the angles of the image blocks; clustering the image blocks according to the angles, constructing test low-resolution image clusters at different angles, iteratively utilizing a non-local prior constraint regularization model and the trained multi-scale progressive convolutional neural networkto iteratively optimize the test low-resolution image clusters at different angles in sequence, and reconstructing an optimal high-resolution image.

Description

Technical field [0001] This application relates to the field of image processing technology, and in particular to an image reconstruction method based on iterative convolutional neural networks. Background technique [0002] In the imaging process of digital image generation, transmission, and recording, high-frequency details are often lost, resulting in low-resolution images. Low-resolution images greatly affect the smooth implementation of various computer vision tasks, such as face recognition in natural scenes, automatic driving, and scene text recognition. Single-image super-resolution reconstruction technology aims to reconstruct a high-resolution image from a single low-resolution image. [0003] So far, three types of image super-resolution reconstruction methods have been proposed. The first type is the typical reconstruction method of nearest neighbor interpolation, bilinear interpolation and bicubic interpolation, which is fast but difficult to reconstruct clear large...

AI Technical Summary

Problems solved by technology

[0005] This application provides an image reconstruction method based on iterative convolutional neural network to solve the technical problem that the existing technology cannot reconstruct the correct image structure

Images