Scholarlink

<p>In today’s visually saturated media landscape, television genres have evolved significantly, embracing hybrid, dynamic, and aesthetically complex patterns. This evolution has transformed traditional genre classification systems, which rely on manual tagging and fixed metadata structures, making them inefficient and often inconsistent. As storytelling formats blend elements from documentaries, fiction, and cinematic production, the lines between genres blur, posing significant challenges for conventional categorization techniques. Addressing this complexity requires intelligent, automated systems capable of detecting visual patterns and semantic cues. In response to these demands, this study investigates the application of advanced vision transformer-based deep learning techniques to classify television content by genre with greater accuracy and adaptability, focusing on the pyramid vision transformer (PVT) architecture. PVT incorporates hierarchical feature extraction with multiscale attention mechanisms, allowing it to capture intricate spatial and semantic details in visual media, to achieve high classification accuracy while significantly reducing computational complexity. Its progressive spatial reduction attention (PSRA) reduces token complexity while retaining context, making PVT more efficient than flat ViTs. By uniting multiscale feature learning with global attention, PVT effectively models hybrid genres where subtle differences in lighting, actor placement, or scene composition are critical. Additionally, explainability through Grad-CAM enhances transparency and practical applicability. Empirical analysis is based on an available dataset comprising 4,475 images across four major TV genres, ensuring a broad representation of content styles. The PVT model’s performance was evaluated against two benchmarks: a conventional convolutional neural network (CNN) and a transfer learning approach using NASNetMobile. The results show that PVT achieves a classification accuracy of 97%, significantly outperforming CNN and NASNetMobile, highlighting its ability to detect visual features such as color gradients, scene composition, and actor placement that are often indicative of genre classification.</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

A Pyramid Vision Transformer–based Framework for Television Video Genre Classification in Multimedia Analysis

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Fawaz Khaled Alarfaj,
  • Hikmat Ullah Khan

摘要

In today’s visually saturated media landscape, television genres have evolved significantly, embracing hybrid, dynamic, and aesthetically complex patterns. This evolution has transformed traditional genre classification systems, which rely on manual tagging and fixed metadata structures, making them inefficient and often inconsistent. As storytelling formats blend elements from documentaries, fiction, and cinematic production, the lines between genres blur, posing significant challenges for conventional categorization techniques. Addressing this complexity requires intelligent, automated systems capable of detecting visual patterns and semantic cues. In response to these demands, this study investigates the application of advanced vision transformer-based deep learning techniques to classify television content by genre with greater accuracy and adaptability, focusing on the pyramid vision transformer (PVT) architecture. PVT incorporates hierarchical feature extraction with multiscale attention mechanisms, allowing it to capture intricate spatial and semantic details in visual media, to achieve high classification accuracy while significantly reducing computational complexity. Its progressive spatial reduction attention (PSRA) reduces token complexity while retaining context, making PVT more efficient than flat ViTs. By uniting multiscale feature learning with global attention, PVT effectively models hybrid genres where subtle differences in lighting, actor placement, or scene composition are critical. Additionally, explainability through Grad-CAM enhances transparency and practical applicability. Empirical analysis is based on an available dataset comprising 4,475 images across four major TV genres, ensuring a broad representation of content styles. The PVT model’s performance was evaluated against two benchmarks: a conventional convolutional neural network (CNN) and a transfer learning approach using NASNetMobile. The results show that PVT achieves a classification accuracy of 97%, significantly outperforming CNN and NASNetMobile, highlighting its ability to detect visual features such as color gradients, scene composition, and actor placement that are often indicative of genre classification.