| Human activity recognition has been a hot research direction in the field of computer vision.The research content of this thesis is based on the recognition of human visual behavior in images.Most of the previous research on behavior recognition mainly focuses on video data,the research content of this thesis makes up for the deficiency of the research on human behavior recognition in images.The main research contents of the thesis include:1.Definition of visual behavior,collection and collation of experimental data.This thesis defines 40 types of visual behavior by referring to the visual behavior categories of video data,verb semantic network library in natural language and common visual behavior in daily life.And make full use of ImageNet,MSCOCO,ImageCLEF,PascalVOC and other large image databases,as well as "Baidu Pictures" and "Google Pictures" search engines to collect the experimental data needed in this thesis,finally,we collected 40 categories,15355 pictures with human visual behavior.2.Design and implementation of image activity recognition model.In this thesis,we first extract image features by convolutional neural network,and then complete the final classification task by traditional classification algorithm.Four hybrid image behavior recognition models are implemented in this thesis: AlexNet + SVM(Support Vector Machine,abbreviated as SVM),AlexNet+RF(Random Forest,abbreviated as RF),AlexNet+NB(Naive Bayesian,abbreviated as NB),AlexNet+DT(Decision Tree,abbreviated as DT),and an AlexNet image activity recognition model.3.Analysis and comparison of experimental results.The thesis analyzes and compares the experimental results of four hybrid models and a single neural network model in detail,and verifies the effectiveness of the proposed image behavior recognition hybrid model.The research results of this thesis show that compared with a single CNN model,the hybrid model is easier to achieve higher accuracy and is more suitable for image activity recognition research,moreover,the experimental data collected in this thesis make up for the shortcomings of the relevant experimental data in this field. |
