Scholarlink

<p>With the rapid growth of digital media, advertisements have become an essential part of video content across various platforms. Effective advertisement classification is crucial for applications such as content moderation, targeted advertising, and video indexing. However, accurately classifying advertisements remains a challenge due to variations in content, background noise, and diverse visual styles<b>.</b> In this paper, a hybrid approach to video advertisement classification is proposed, which integrates an improved segmentation process and an optimized classification algorithm (ImpSN-GN). The system starts with video preprocessing, converting input videos into frames and using a median filter to eliminate noise. A better Mask R-CNN model with scaled dot-product and multi-head attention mechanisms is utilized for segmentation, which improves accuracy and minimizes gradient instability. Discriminative features such as altered motion estimation features, multi-texton features, shape features, and statistical features are derived from the segmented frames. They are then classified with the ImpSN-GN model, which is the combination of Improved ShuffleNet (ImpSN) and GhostNet (GN).</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

ImpShuffleNet-GhostNet approach for classification of advertisement from video

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • P. Santhi,
  • B. Lanitha

摘要

With the rapid growth of digital media, advertisements have become an essential part of video content across various platforms. Effective advertisement classification is crucial for applications such as content moderation, targeted advertising, and video indexing. However, accurately classifying advertisements remains a challenge due to variations in content, background noise, and diverse visual styles. In this paper, a hybrid approach to video advertisement classification is proposed, which integrates an improved segmentation process and an optimized classification algorithm (ImpSN-GN). The system starts with video preprocessing, converting input videos into frames and using a median filter to eliminate noise. A better Mask R-CNN model with scaled dot-product and multi-head attention mechanisms is utilized for segmentation, which improves accuracy and minimizes gradient instability. Discriminative features such as altered motion estimation features, multi-texton features, shape features, and statistical features are derived from the segmented frames. They are then classified with the ImpSN-GN model, which is the combination of Improved ShuffleNet (ImpSN) and GhostNet (GN).