基于深度Q网络的红外图像去雾研究
首发时间:2024-06-19
摘要:近年来,随着我国红外热成像技术的快速发展和其在多个领域的广泛应用,对红外图像的质量和清晰度提出了更高要求。然而,全球气候环境的变化和空气污染加重,雾霾天气出现日益增多,严重影响了红外图像的质量和清晰度。随着计算机技术的发展与普及,基于深度强化学习的计算机图像处理技术在红外图像处理中也扮演着越来越重要的角色。因此,如何利用深度强化学习技术有效减弱红外图像中的雾霾,提高红外图像的质量和清晰度,具有重要的理论意义和实际应用价值。针对红外图像的特性,提出使用深度Q网络算法实现红外图像去雾。将深度Q网络算法、图像先验算法和图像增强算法相结合,利用强化学习的智能体与环境的交互,实现去雾效果的主动判断,得到了基于深度强化学习的红外图像去雾新方法,有效减弱了不同浓度的雾霾对红外图像的影响。在深度Q网络模型中设计了去雾框架、状态空间、动作空间和奖励函数。为了进一步提升网络模型的性能和稳定性,引入了优先经验回放机制,使得模型在训练中能更高效地利用有价值的经验,从而加速学习进程并提高去雾效果。
For information in English, please click here
Research on infrared image dehazing based on deep Q-network
Abstract:In recent years, with the rapid development of infrared thermal imaging technology in China and its widespread application in multiple fields, higher requirements have been put forward for the quality and clarity of infrared images. However, the changes in global climate and environment, as well as the worsening of air pollution, have led to an increasing occurrence of haze weather, seriously affecting the quality and clarity of infrared images. With the development and popularization of computer technology, computer image processing technology based on deep reinforcement learning is also playing an increasingly important role in infrared image processing. Therefore, how to effectively reduce haze in infrared images and improve the quality and clarity of infrared images using deep reinforcement learning technology has important theoretical significance and practical application value. This thesis focuses on infrared image dehazing technology, and the main research content is as follows: A deep reinforcement learning-based method for infrared image dehazing, using the deep Q-network algorithm, is proposed. By combining deep Q-network algorithm, image prior algorithm, and image enhancement algorithm, and utilizing the interaction between reinforcement learning agents and the environment, a new method for infrared image dehazing based on deep reinforcement learning is achieved, which effectively reduces the impact of different concentrations of haze on infrared images. A dehazing framework, state space, action space, and reward function were designed in the deep Q-network model. In order to further improve the performance and stability of the network model, a priority experience replay mechanism has been introduced, which enables the model to more efficiently utilize valuable experience during training, thereby accelerating the learning process and improving the dehazing effect
Keywords: infrared image image dehazing deep reinforcement learning Dark channel prio
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
基于深度Q网络的红外图像去雾研究
评论
全部评论