Scholarlink

<p>Image matting plays a crucial role in image synthesis and editing, particularly in applications such as live broadcasting, film production, and micro-classes. Traditional algorithms, while successful in natural image matting, struggle with green screen images due to issues like color overflow and complexity in local areas. To address these challenges, we propose a novel green screen matting network model built upon the Background Matting framework. Our model integrates group normalized stability training, incorporates an Atrous Spatial Pyramid Pooling (ASPP) module and Deformable Part Model (DPN) module, and introduces perceptual loss to establish pixel correlations. Furthermore, we construct a new green screen dataset by fusing composite and real images from diverse scenes, enhancing the model’s generalization ability. Experimental results demonstrate that our approach outperforms existing models in handling complex edges and green screen images, particularly in scenarios with uneven lighting or shadows. Our lightweight model, with fewer parameters, is also more suitable for industrial applications. The code is available at a Github repository: <a href="https://github.com/dongji00/EGSM">https://github.com/dongji00/EGSM</a>.</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 green screen matting with group normalization and perceptual loss for color overflow and complex edges

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Bo Chen,
  • Chenyu Zhou,
  • Xiaoli Sun

摘要

Image matting plays a crucial role in image synthesis and editing, particularly in applications such as live broadcasting, film production, and micro-classes. Traditional algorithms, while successful in natural image matting, struggle with green screen images due to issues like color overflow and complexity in local areas. To address these challenges, we propose a novel green screen matting network model built upon the Background Matting framework. Our model integrates group normalized stability training, incorporates an Atrous Spatial Pyramid Pooling (ASPP) module and Deformable Part Model (DPN) module, and introduces perceptual loss to establish pixel correlations. Furthermore, we construct a new green screen dataset by fusing composite and real images from diverse scenes, enhancing the model’s generalization ability. Experimental results demonstrate that our approach outperforms existing models in handling complex edges and green screen images, particularly in scenarios with uneven lighting or shadows. Our lightweight model, with fewer parameters, is also more suitable for industrial applications. The code is available at a Github repository: https://github.com/dongji00/EGSM.