Semantic Based Video Lecture Summarization by Using Deep Networks

Abstract

To overcome the shortcomings of current approaches, which mostly concentrate on controlled offline recordings, this study offers a novel AI-driven framework for semantic summarisation of online lecture videos. Both conciseness and educational fidelity are improved by the suggested four-stage methodology, which consists of fuzzy-based deep semantic summarisation, multimodal feature fusion, audio denoising, and video preprocessing. Through adaptive contrast enhancement, noise reduction, and Lanczos resampling, the LR + BM3D + CLAHE pipeline enhances video quality during preprocessing, reaching up to 0.945 accuracy and 0.942 precision. Adaptive levellers, spectrum subtraction, and high-pass filtering are used by the audio module to improve clarity, lowering the MSE from 0.243 to 0.092 and raising the SNR from 0.355 to 0.434 dB. We extract visual and audio features using HBBEA-DRN and YCbCr segmentation to produce summaries that retain over 90% informational density with 55 58% time savings. To achieve precision, recall, and F1 gt 0.90 on unseen movies, the fuzzy-based Deep Conceptual Knowledge Network (DCKN) combines semantic clues utilising interpretable fuzzy rules. According to comparative analysis, our method outperforms DASP (87.1%), MASN (85.8%), and FCN-LectureNet (86.4%) with an accuracy of 92.3%. The system sets a new standard for intelligent lecture summarisation in online learning settings with its outstanding scalability, resilience, and semantic purity. newline