Based on the RGB images extracted through P3D Resnet and the artificial intelligence model learned based on Opitcal flow images, this paper proposes a detection model using violent image data from public data, and applies it to the infant abuse detection system through transfer learning to prepare measures to prevent and commercialize child abuse.
The dataset was first constructed for the research and algorithm composition of the paper. The Preudo 3D Conv network model was constructed using the public data of violent images provided by AI Hub to prepare for the analysis of the accuracy of violence detection.
AI Hub is an integrated platform that anyone can use and participate in by supporting AI infrastructure (AI data, AI SW API, and computing resources) necessary for AI technology, products, and services development, and has various data such as images and videos.
For this study, image data on Assault among the datasets of "abnormal behavior CCTV images" were collected, and inspection work was conducted to ensure the accuracy and research validity of the data.
Accordingly, 965 video data were secured, and 30 additional types of child abuse video data were secured to confirm transfer learning to conduct a study.
The technology development process of this study collected sufficient data to analyze and learn the data, then conducted an experiment on the Resnet50/Resnet101-based Pseudo 3D Conv network model, and then conducted a PCA conversion and learning experiment on the Resnet101-based Pseudo 3D Conv network model.
As a result, it is judged that it is more advantageous to search for an object's motion by using only the motion information of an object in an image than to extract features from unnecessary information provided by RGB images.
In addition, we present the possibility that further supplementing algorithms for preprocessing images, such as PCA, can improve the performance of the model and thus reduce the weight of the massive configured model.