Automatic parental guide ratings for short movies

Chai, Zi Xu (2021) Automatic parental guide ratings for short movies. Final Year Project, UTAR.

Preview

PDF Download (2930Kb) | Preview

Abstract

Video description is helpful for automatic movie ratings and annotating parental guides. However, human-annotated ratings are somewhat ambiguous depending on the types of movies and demographics. This project proposes a Machine-learning (ML) pipeline to generate a parental rating for short movies automatically. The ML pipeline infers and resolves various entities from 5 custom trained ML models trained using a corresponding public dataset. These ML models include nudity scene detection, violent scene detection, profanity scene detection, alcohol & drugs detection. A nudity detection scene is trained using YOLOv4 to detect possible scenes exposing private parts and genitals. Meanwhile, violent scene detection is trained using custom RNN-LSTM to detect possible fighting and gore scenes. Next, the profanity detection uses Google Text-to-Speech API to transcribe audio before feeding it into a custom better-profanity library. Lastly, the alcohol & drug models are trained using features extracted from VGG-16 then fed into a one-class CNN classifier. The experimental result showed that the proposed automatic rating is highly accurate when compared to manually annotated ratings.

Actions (login required)