Scholarlink

<p>Due to the limitations of imaging equipment, only some clear information in the real scene can be obtained through a single type of sensor, and other parts are often blurred. In light of the imaging mechanism of the sensor, multi-focus images are frequently generated as a result of the depth of field information present in the image scene. Images with different focus points can be obtained by multi-sensors with different Settings, and multi-focus images can be synthesized into clear images by multi-focus image fusion. Generating a decision image with a precise boundary is the key to multi-focus image fusion. We employed the dataset-creation approach utilized in our prior work. The depth of field estimation model was applied to acquire the depth of field information. Following this, an artificial multi-focus image dataset was created in accordance with the given depth of field information. Our double-branch encoder combines Transformer block and invertible neural network (INN), and uses Swin-Transformer block to design the corresponding decoder, which combines to get a U-shaped network. Then, the depth of field information obtained from the depth estimation model is used to design the loss function. In this work, we put forward a novel double-branch encoder U-net which is capable of effectively extracting the global information and details of multi-focus image pairs, and can generate clear fused images during the fusion process. Ultimately, we illustrate the preeminence of our approach via both qualitative and quantitative assessments. Our code is available at <a href="https://github.com/hyukshao/MFIF-TCDU-Net.git">https://github.com/hyukshao/MFIF-TCDU-Net.git</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

Multi-focus image fusion based on double branch encoder and depth information learning

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Xin Jin,
  • Xinfeng Shao,
  • Shiwei Wang,
  • Chengyi Pan,
  • Shaowen Yao,
  • Qian Jiang

摘要

Due to the limitations of imaging equipment, only some clear information in the real scene can be obtained through a single type of sensor, and other parts are often blurred. In light of the imaging mechanism of the sensor, multi-focus images are frequently generated as a result of the depth of field information present in the image scene. Images with different focus points can be obtained by multi-sensors with different Settings, and multi-focus images can be synthesized into clear images by multi-focus image fusion. Generating a decision image with a precise boundary is the key to multi-focus image fusion. We employed the dataset-creation approach utilized in our prior work. The depth of field estimation model was applied to acquire the depth of field information. Following this, an artificial multi-focus image dataset was created in accordance with the given depth of field information. Our double-branch encoder combines Transformer block and invertible neural network (INN), and uses Swin-Transformer block to design the corresponding decoder, which combines to get a U-shaped network. Then, the depth of field information obtained from the depth estimation model is used to design the loss function. In this work, we put forward a novel double-branch encoder U-net which is capable of effectively extracting the global information and details of multi-focus image pairs, and can generate clear fused images during the fusion process. Ultimately, we illustrate the preeminence of our approach via both qualitative and quantitative assessments. Our code is available at https://github.com/hyukshao/MFIF-TCDU-Net.git.