CycleGAN-based image training network structure ArcGAN and method

Abstract

The invention discloses a CycleGAN-based image training network structure ArcGAN and a method, the network structure ArcGAN is composed of a generator and double discriminators, and the double discriminators comprise a rough discriminator and a fine discriminator, wherein the encoder comprises an input layer and three down-sampling convolution layers, and each down-sampling convolution layer is connected with two flat convolution layers with the same structure as the input layer; a converter comprises five dense convolution blocks without pooling layers, each dense convolution block comprisesfive dense convolution layers with bottleneck layers, and a compression layer is arranged between the blocks, wherein the decoder comprises three upsampling deconvolution layers and an output layer, and each upsampling deconvolution layer is connected with two flat convolution layers with the same structure as the input layer; and each layer of downsampling in the encoder and the upsampling in thedecoder corresponding to the layer of downsampling are in replicated connection. A rough discriminator is used for processing high-level visual information and consists of six down-sampling layers and an output layer; the calculation loss of the fine discriminator is combined with the calculation loss of the rough discriminator, and the fine discriminator and the rough discriminator jointly complete confrontation consistency training.

Description

technical field [0001] The present invention mainly relates to deep learning and image processing field, relate in particular to a kind of network structure ArcGAN based on the picture training of CycleGAN and the method for line building automatic coloring. Background technique [0002] Since the mid-1990s, there has been a large body of research exploring how to automatically transform images into synthetic artwork with a particular style. Pioneering work by Gatys et al. demonstrates the power of convolutional neural networks (CNNs) to create artistic imagery by separating and recombining image content and style [2] . The process of using CNN to render content images with different styles is called Neural Style Transfer (NST). Since then, NST has become a hot topic in academia and industry, it is receiving more and more attention, and various methods to improve or extend the original NST algorithm have been proposed. [0003] Deep Neural Networks are the most used, most...

AI Technical Summary

Problems solved by technology

[0020] The purpose of the present invention is to overcome the deficiencies in the prior art, and provide a network structure ArcGAN and method based on CycleGAN image training. The style migration of buildings faces three problems: the existing tools for generating building drawings cannot achieve style The limited available training set and the existing deep network structure cannot be well applied to the data set of the present invention to produce ideal results

The invention can solve the problem of style migration from line architectural drawings to colored architectural drawings of different styles

Images