Scholarlink

<p>Recognizing human actions in aerial videos is challenging due to reduced resolution and blurry appearance of humans. We propose GAN-SE, a two-module system to address these issues. The first module employs a novel super-resolution GAN to enhance the low-resolution images of detected humans. By generating high-resolution images, our system enhances the visual quality of the detections, thereby improving action recognition accuracy. The second module combines a squeeze-and-excitation network with ResNeXt101 to recalibrate feature responses, focusing on the most relevant information. This leads to superior feature representation and more accurate predictions. To evaluate the performance of GAN-SE, extensive experiments were conducted on three challenging datasets: Aeriform in-action, UCF-ARG, and Okutama-Action. The results demonstrate the effectiveness of our system, achieving an accuracy of 80.78, 97.36, and 77.50% on the respective datasets. These results outperform the state-of-the-art methods, reaffirming the superiority of GAN-SE in aerial human action recognition.</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

Enhancing aerial human action recognition through GAN-boosted ResNeXt architecture with squeeze-and-excitation network

  • 期刊论文

  • 发表日期：

    2025年12月11日

查看全文
  • Surbhi Kapoor,
  • Akashdeep Sharma,
  • Amandeep Verma

摘要

Recognizing human actions in aerial videos is challenging due to reduced resolution and blurry appearance of humans. We propose GAN-SE, a two-module system to address these issues. The first module employs a novel super-resolution GAN to enhance the low-resolution images of detected humans. By generating high-resolution images, our system enhances the visual quality of the detections, thereby improving action recognition accuracy. The second module combines a squeeze-and-excitation network with ResNeXt101 to recalibrate feature responses, focusing on the most relevant information. This leads to superior feature representation and more accurate predictions. To evaluate the performance of GAN-SE, extensive experiments were conducted on three challenging datasets: Aeriform in-action, UCF-ARG, and Okutama-Action. The results demonstrate the effectiveness of our system, achieving an accuracy of 80.78, 97.36, and 77.50% on the respective datasets. These results outperform the state-of-the-art methods, reaffirming the superiority of GAN-SE in aerial human action recognition.