Scholarlink

<p>This paper presents a framework for multi-focus image fusion based on reinforcement learning using Deep Q-Network (DQN). The proposed method aims to merge multiple images focused at different depths into a single high-quality image that retains sharp details from both sources. At the core of the model is a Deep Q-Network (DQN), which enables the fusion process by selecting the optimal image patches from two input images based on learned quality metrics. The model is trained using reinforcement learning, where the agent receives rewards based on sharpness and Structural Similarity Index (SSIM). This helps to select the best patches for fusion. Convolutional layers in the DQN architecture extract relevant features from the images while fully connected layers make the final patch selection decision. Extensive experimentation demonstrates the effectiveness of the approach with performance evaluated using objective quality measures such as PSNR, SSIM, Entropy, Edge Intensity and Cross Entropy. The results show that the model outperforms traditional fusion methods in terms of image clarity, sharpness and detail retention making it a promising solution for various applications.</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

DeepQFusion: A reinforcement learning-based framework for multi-focus image fusion using deep Q-Networks

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • M. S Santhi Krishna,
  • D Harikrishnan,
  • C R Shiyas

摘要

This paper presents a framework for multi-focus image fusion based on reinforcement learning using Deep Q-Network (DQN). The proposed method aims to merge multiple images focused at different depths into a single high-quality image that retains sharp details from both sources. At the core of the model is a Deep Q-Network (DQN), which enables the fusion process by selecting the optimal image patches from two input images based on learned quality metrics. The model is trained using reinforcement learning, where the agent receives rewards based on sharpness and Structural Similarity Index (SSIM). This helps to select the best patches for fusion. Convolutional layers in the DQN architecture extract relevant features from the images while fully connected layers make the final patch selection decision. Extensive experimentation demonstrates the effectiveness of the approach with performance evaluated using objective quality measures such as PSNR, SSIM, Entropy, Edge Intensity and Cross Entropy. The results show that the model outperforms traditional fusion methods in terms of image clarity, sharpness and detail retention making it a promising solution for various applications.