Texture Synthesis Method of Arbitrary Size Samples Based on Convolutional Neural Network

Abstract

The invention discloses a random-size sample texture synthesis method based on a convolutional neural network. The method includes the steps: (1) inputting a 512*512 texture image to be processed; (2)building and training the convolutional neural network; (3) driving the texture image to be processed; (4) forming a synthesis image matrix of the texture image to be processed; (5) forming a synthesis image of the texture image to be processed. According to the method, the convolutional neural network is led into the field of texture image synthesis, and the method overcomes the shortcomings that local optimum is easily caused by optimal matching, and random-size texture images cannot be synthesized in the prior art. According to the method, profiles of the synthesized texture images are clearer, realer and more natural.

Description

technical field [0001] The invention belongs to the technical field of image processing, and further relates to a method for synthesizing sample textures of arbitrary sizes based on convolutional neural networks in the technical field of natural image processing. The invention aims at all irregular natural texture images, adopts convolutional neural network, and can be used for synthesis of texture images of any size. Background technique [0002] Currently, texture synthesis has become a very important research topic in the field of image processing technology. According to the different basic units selected for sample texture synthesis, the texture synthesis methods under the MRF (Markov Random Field) model can basically be divided into two categories: one is pixel-based texture synthesis; the other is block-based Texture compositing. The two types of methods have their own strengths. The pixel-based method is good at capturing the local details of the texture, but it is...

AI Technical Summary

Problems solved by technology

However, the disadvantage of this method is that the best matching used will easily lead to a small number of acquired features, and the result is a local optimum. At the same time, the matching between pixels is prone to errors, making the synthesized texture image easy to obtain. Blurred and unable to composite texture images of arbitrary size

However, the disadvantage of this method is that using the Euclidean distance on the pixel domain can easily cause the synthesized texture to contain many broken structures. For low-resolution input images, the algorithm cannot complete texture synthesis.

Images