Research On Recognition Model Of Stick Figure Based On Deep Learning

Stick figure has been one of the ways of human daily communication since the primitive society.Nowadays,with the popularization and development of mobile electronic devices such as mobile phones,computers,tablets and touch screen technology,image information has gradually become important in human daily communication.Status and research related to stick figures have gradually become a hot spot in the field of artificial intelligence.Early stick figure recognition was based on traditional image recognition methods based on artificial feature zones,such as Direction Gradient Histogram(HOG),Scale Invariant Feature Transform(SIFT),etc.,but these feature methods were mainly designed for natural images.Compared with natural images,strokes have the characteristics of abstractness,randomness,and lack of color and texture information,which makes these image recognition methods not completely used for simple stroke recognition.At the same time,manual feature extraction is not only timeconsuming and laborious,but also has high requirements for feature extraction experience.The rise of deep learning is a good solution to the problem of manually extracting features.It is a computer that simulates the network structure of biological receiving and learning information to enable the computer to automatically analyze and learn image information.Higher-level,more abstract feature information,these characteristics determine the superiority of deep neural networks in the field of image recognition.This article will use deep neural networks to conduct research on stick figure recognition.Although theoretically,the deeper the network level,the more abstract features can be learned,but at the same time the shallow feature information may be filtered out,resulting in recognition accuracy decline.Too deep network levels may cause gradient dispersion,network performance degradation,and other issues.This article will address this problem by balancing network depth and width,combining residual neural networks and Inception structures,and adding features that can effectively avoid overfitting,Strategies such as batch normalization and Dropout layer for problems such as gradient dispersion have constructed an improved multi-channel residual neural network.By analyzing the impact of each hyperparameter on the recognition rate of the test set,a set of hyperparameters was finally determined to train a model with high recognition rate and good network performance,and compared with other neural networkbased recognition methods and traditional image recognition methods To verify the effectiveness of the proposed model in the field of stick figure recognition.On the other hand,considering that insufficient training may lead to overfitting problems,two data enhancement schemes are also proposed in the article,and it is verified that these schemes can improve the network recognition performance to a certain extent.Finally,based on the B / S architecture and MVC design ideas,the deep neural network model proposed in the paper is deployed to the Java platform,and a simple stick figure recognition system is designed so that it can be applied on the web side.