Image stylizing method capable of being operated in real time

Abstract

The invention discloses an image stylizing method capable of being operated in real time. The method comprises the following steps: forming the encoding part of a generation network by using a fire module proposed in Squeeze-Net and forming a decoding part of the generation network by using a de-convolution layer; forming a discrimination network by using conv1_1 to conv5_1 convolution layers in the VGG19; performing semi-supervised learning by constructing a generation confrontational network; subjecting the de-convolution layer to flattened decomposition, dimensionality reduction, and elimination of redundancy by using a convolution kernel low-rank decomposition method on the premise that the picture synthesis capability of the generation network is not changed, and subjecting the whole generation network to further acceleration and compression by using a network pruning and parameter quantification method. The image stylizing method applies the Fire module and the Flat-Deconv module to the generation network for the first time, and integrates multi-scale information during encoding and decoding. Thus, the generation network is made lightweight when a certain synthesis capability of the generation network is guaranteed, so as to be more suitable for mobile terminals.

Description

technical field [0001] The invention relates to the algorithm field of image processing and deep learning, in particular to an image stylization method capable of real-time operation. Background technique [0002] Image editing applications have long been numerous, but most of them use simple filters to change the color, light, contrast, etc. of pictures. Even some image special effects like oil paintings and sketches are artificially designed features and strokes to render pictures. Not only the algorithm development cycle is long, but also it takes a lot of time to run on the terminal. [0003] An oral article by CVPR "Image Style Transfer Using Convolutional NeuralNetworks" has attracted people's attention in the past six months. This is a method of using deep learning algorithms for image stylization learning. The core idea behind it is to use deep convolutional neural networks The network can separate content features and style features at different scales, making it e...

AI Technical Summary

Problems solved by technology

This algorithm makes the design of image special effects free from human intervention to a certain extent. However, the entire optimization process requires a large number of floating point operations. more serious

Images