Abstract:

With the explosive growth of user-generated content on platforms like YouTube, the presence of inappropriate or harmful content has become a critical concern for platform providers, regulators, and viewers alike. Traditional moderation techniques, including manual review and basic keyword filtering, are not scalable or effective enough to handle the vast and continuously growing volume of video data. In this context, this project proposes a deep learning-based approach for detecting and classifying inappropriate content in YouTube videos, focusing on categories such as hate speech, nudity, violence, offensive language, and misinformation. The proposed system integrates Natural Language Processing (NLP) and Computer Vision (CV) techniques using deep learning models. For audio and textual analysis (e.g., captions, comments, and transcripts), advanced NLP models such as BERT are employed to understand context and semantics. For visual content, Convolutional Neural Networks (CNNs) and pre-trained models like ResNet or EfficientNet are used to analyze video frames for visual cues of inappropriate material. Additionally, Recurrent Neural Networks (RNNs) or LSTM layers are used to understand the temporal sequence of actions or dialogues in videos. By combining multimodal data—text, audio, and video—the system achieves a higher accuracy rate in detecting nuanced and context-dependent inappropriate content. The model is trained and validated on a labeled dataset of YouTube videos, and its performance is evaluated using precision, recall, F1-score, and accuracy metrics. The ultimate goal of this research is to support safer online environments by enabling automated, scalable, and intelligent content moderation for video-sharing platforms.

Description:

.