Scholarlink

<p>In video surveillance, distortions such as blurred focus, haze, motion blur, and uneven illumination, as well as their simultaneous combinations, adversely affect video quality and hinder the effectiveness of various high-level tasks, such as object detection and identification, visual tracking, abnormal event detection, and scene understanding. Existing deep learning methods struggle with generalization and handling of coexisting distortions in real-world scenarios. To address these challenges, we propose a new architecture that combines a vision transformer scheme and attention mechanisms modules for identifying and classifying multiple distortions in surveillance videos. Specifically, we adapt and fine-tune a pretrained ViT-B/16 model, incorporating specialized classifier heads to handle both single and multiple distortions in the VSQuAD dataset. Our method leverages a vision transformer for its robust feature extraction capabilities and attention mechanisms, enabling it to distinguish subtle and globally dispersed distortion patterns. To enhance the model’s robustness against class imbalance, we employ weighted cross-entropy loss, data augmentation, and adaptive learning rate scheduling. The method achieves 95.8% test accuracy on single distortions only, 95.1% on multiple distortions only, and 91.1% on a combination of both. Our method demonstrates superior performance in terms of distortion classification on the VSQuAD dataset compared to the state-of-the-art using the same dataset. These results demonstrate that our method outperforms existing approaches and has applications in real-time surveillance, such as remote surgery and video-based transportation systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LViT-VDC: A Lightweight Vision Transformer-Based Architecture for Video Distortion Classification

  • 期刊论文

  • 发表日期：

    2026年5月16日

查看全文
  • Hammad Hassan Gillani,
  • Muhammad Ali Qureshi,
  • Azeddine Beghdadi,
  • Azzedine Zerguine

摘要

In video surveillance, distortions such as blurred focus, haze, motion blur, and uneven illumination, as well as their simultaneous combinations, adversely affect video quality and hinder the effectiveness of various high-level tasks, such as object detection and identification, visual tracking, abnormal event detection, and scene understanding. Existing deep learning methods struggle with generalization and handling of coexisting distortions in real-world scenarios. To address these challenges, we propose a new architecture that combines a vision transformer scheme and attention mechanisms modules for identifying and classifying multiple distortions in surveillance videos. Specifically, we adapt and fine-tune a pretrained ViT-B/16 model, incorporating specialized classifier heads to handle both single and multiple distortions in the VSQuAD dataset. Our method leverages a vision transformer for its robust feature extraction capabilities and attention mechanisms, enabling it to distinguish subtle and globally dispersed distortion patterns. To enhance the model’s robustness against class imbalance, we employ weighted cross-entropy loss, data augmentation, and adaptive learning rate scheduling. The method achieves 95.8% test accuracy on single distortions only, 95.1% on multiple distortions only, and 91.1% on a combination of both. Our method demonstrates superior performance in terms of distortion classification on the VSQuAD dataset compared to the state-of-the-art using the same dataset. These results demonstrate that our method outperforms existing approaches and has applications in real-time surveillance, such as remote surgery and video-based transportation systems.