Semantic Classification And Application For Cartoon Images Based On Machine Learning

Image classification based on machine learning refers to using machine learning to generate a neural network,which can give the category of the corresponding picture.This method is faster and more efficient than the traditional image recognition method,and is more and more applied in the actual application system.However,there is almost no classification systems for hand-drawn cartoon images based on machine learning.This paper studies and implements how to construct such systems and their possible applications.At present,most of the data sets that can be used to train the image classification network are pictures collected in the real world,but the classification ability of the neural network is inseparable from the data set used by the network.The experimental results show that if real pictures are used to train the neural network,the network is unable to classify cartoon pictures.Not only that,the existing advanced classification neural networks are also optimized for real image datasets,and there is no neural network structure tailored for cartoon design.Aiming at the current situation of rare cartoon data sets and few classification neural networks for cartoon painting,this paper proposes a semantic classification system for hand-drawn cartoons based on machine learning.The research results of this paper are as follows:(1)A data set generation system is implemented.We built our benchmark dataset from 4,000 images of 11 different categories collected from the Internet and expanded our dataset to 10,000 images using three methods.These methods include:First,using a custom cartoon shader,a special lighting model combined with an edge detection shader,thus the rendered 3D model has a hand-drawn style,and finally a multi-angle capture method applied for generating cartoon images;Secondly,using the existing cartoon painting modeling application,the 2D image is modeled into a 3D model,and then the hand-drawn style is rendered by a custom shader.Finally,the multi-angle capture method applied;Thirdly,using the hand-drawn stylized filter,the original cartoon image is transformed into a variety of different cartoon styles.The image stylization methods such as color pencil style and crayon style are implemented,and the data diversity of the data set is further improved.(2)A classification neural network system is implemented.The neural network architecture used in this paper has three main strategies.By using these three technologies,the classification accuracy of the system is 5% higher than the state-of-the-art network.These strategies include:First,we introduce an input unified stylization strategy.The main idea of the strategy is to improve the classification accuracy by reducing the complexity of the image without reducing the picture information by preprocessing the input picture of the network structure.Secondly,the feature inserted network.The neural network structure inserts global features into specific positions of the neural network.Because the background complexity of the cartoon drawing is low,and the foreground is mostly characterized by large color blocks,the statistical information such as color histograms is inserted,which can improve the classification ability of the network;Third,network plus network.The neural network architecture firstly separates and extracts network themes by pre-training multiple single networks,superimposing feature layers of specific locations as a new hybrid network,and finally training into a new network to obtain a more powerful feature layer which further enhances the system's classification capabilities.(3)This paper also proposes the application extension based on the classification result.The sound and skeleton animation information is added in cartoon painting modeling,and the skeleton extraction method combining semantic information is implemented,which provides the forward direction of our future work.