International Journal of Computer Science and Technology

International Journal of Computer Science and Technology | Vol. 4, No. 1, January 2023 | pp. 1–8

DOI: 10.46882/2023/IJCST/000216

Article Type: Original Research Paper

Title: Enhancing Low-Light Video Summarization via Spatial-Temporal Attention Network Models

Names of Authors: Min-Ji Kim¹, David Vance²

Authors’ Affiliations: ¹School of Electronics and Informatics, Seoul National University, Seoul, South Korea; ²Computer Vision Research Systems, Lumina Tech Labs, Denver, USA

Abstract: Processing video security streams recorded in low-light environments is challenging due to high sensor noise and low color contrast. These artifacts cause standard deep learning object trackers and video summarization models to drop critical frames. This paper introduces a specialized spatial-temporal attention model designed to summarize low-light security video footage accurately. The architecture features an internal feature-enhancement layer that amplifies low-contrast details without over-processing high-frequency sensor noise. A recurrent neural layer calculates regional attention maps across consecutive frames to highlight moving subjects. We trained and evaluated the neural system using a comprehensive low-light security dataset. The model achieved a F1-score optimization level of 0.89, outperforming conventional summarization networks by 14.2%. The total video file processing volume was condensed by 65.0%, preserving critical action sequences. This framework provides an efficient method to scale large-scale urban video analytics infrastructure running under challenging environmental conditions.

Keywords: Video Summarization, Spatial-Temporal Attention, Low-Light Enhancement, Computer Vision, Deep Learning, Security Analytics

Manuscript Timeline: Received: June 02, 2022; Revised: July 08, 2022; Accepted: August 01, 2022; Published: January 06, 2023

Subscribe to read the full article: https://internationalscholarsjournals.org/subscribe-to-read