Small data set stick figure generation method based on BPL

Abstract

The invention discloses a small data set sketch generation method based on BPL. The method comprises the steps: learning strokes, stroke positions and stroke position association, generating probability models of various types, and carrying out the implementation of sketches according to different probability models; wherein the probability model of the category T meets the following conditions: eis one example of the category T, I is a specific image of the example e, M is the number of examples corresponding to the category T, P is the probability, and m is the cycle number. According to the small data set stick figure generation method based on the BPL, the image with the good effect can be generated based on the small data set, the current situation that AI drawing can be conducted only through a large amount of original data is changed, efficiency is improved, and cost is reduced; on the basis of realizing the simulation of the original painting style, the template can be recreated according to the original painting style.

Description

technical field [0001] The invention relates to a BPL-based small data set stick figure generation method, which belongs to the stick figure generation field. Background technique [0002] AI painting is a research hotspot in computer vision. With the continuous progress of related work, AI painting has achieved rapid growth and has become a hot topic that everyone pays attention to. From stick figures to paintings, and then to the generation of real pictures, even humans are difficult to distinguish . AI painting refers to a computer program for automatic digital drawing based on machine learning models. Although AI painting is developing rapidly, there are still many severe challenges in AI painting. [0003] For machine learning models, the process of letting AI learn to paint is a model construction and training adjustment of related parameters. In model training, each picture is essentially an m*n pixel matrix; for color images That is, each pixel is composed of RGB t...

AI Technical Summary

Problems solved by technology

Although the principle of AI painting model training is simple, it is not easy to implement

In 2012, Wu Enda and Jeff Dean used Google Brain's tens of thousands of CPUs to train the neural network to generate cat face pictures. During the training process, the two scholars used a total of 10 million cat face pictures. The model training used Three days is a relatively long time, but the final effect is not satisfactory

One reason is that "creation" is a relatively abstract concept. For models, it is difficult to use models to express abstract "styles"

Images