基于深度特征学习的图像恶劣天气识别与增强技术研究

Weather is an important factor that affects the quality of outdoor images. Image captured under bad weather conditions is seriously degraded, which affects the observability of images and the accuracy of feature extraction. There are three main problems in current research of image enhancement. 1) Traditional hand-craft image features cannot be used directly to describe the weather. 2) The existing methods lead to severe information loss on rain removal results, they are time-consuming that cannot use in real-time systems. 3) The existing methods fail to adaptively dehaze on any haze-density images, that can result in excessive removal of the light haze-density image and insufficient removal of the heavy haze-density image. In order to solve the above problems, this project studies the technology of image weather recognition and enhancement based on deep feature learning. The main research contents are as follows. 1) Image weather recognition based on multi-scale deep feature learning and multiple loss functions fusion. This technology aims to improve the recognition accuracy and overcome the limitations of traditional methods on weather categories and scenes. 2) Improving the visual model of rain image, and propose a sparse dark channel guided fully convolutional network to realize the rapid and simultaneous removal of rain and haze. 3) An attribute disentangled generative adversarial network is proposed to generate high-quality haze removal image by controlling the attribute values and improving the loss function. Finally, an image weather recognition and enhancement system will be built to comprehensively evaluate the proposed methods in outdoor applications such as intelligent video surveillance and unmanned driving systems.

天气是影响图像质量的重要因素，恶劣天气下拍摄的图像信息损失严重，影响图像的可观测性及特征提取的准确性。当前针对图像天气识别与增强研究的不足主要有三点：1）传统图像特征无法直接用于区分天气；2）去雨结果信息损失严重、耗时长；3）无法自适应去雾，导致轻雾图像去除过度、重雾图像去除不足。为解决以上问题，本项目拟采用基于深度特征学习的方法，研究图像恶劣天气识别与增强技术，主要研究内容包括：1）基于多尺度特征学习和多损失函数融合的图像天气识别技术，旨在提高识别效果且突破传统方法对可识别天气类别和场景的限制；2）改进雨的视觉模型，提出融合稀疏暗通道的全卷积神经网络，实现图像雨雾快速同步去除；3）提出一种基于残差结构的属性分解对抗神经网络，通过控制属性值、改进损失函数生成清晰的去雾图像。最后，开发一套图像恶劣天气识别与增强原型系统，在智能监控、无人驾驶等实际应用中对本项目研究成果进行充分验证。

恶劣天气是影响图像质量的重要因素，恶劣天气下拍摄的图像通常存在图像噪声大、对比度低、细节信息损失严重等现象，不但影响图像的可观测性，还会影响特征提取的准确性，导致图像分析算法性能下降甚至失效。本项目以面向室外场景的全天候智能监控为应用背景，对任意场景的天气识别和低质图像增强技术进行了深入的研究，取得了如下成果：1）提出一种基于协作群智感知的图像及环境数据采集方法；2）提出一种基于多尺度特征学习的图像天气识别算法；3）提出一种基于多任务联合的低质图像增强算法；4）开发面向室外监控场景的天气识别与增强原型系统。项目组累计发表学术论文9篇（包含SCI期刊论文5篇，会议论文4篇）、授权发明专利1项、申请发明专利2项、培养研究生10名（包含博士生5名，硕士生5名），按期完成项目研究计划所列内容，基本达到预期指标。项目组提出的模型、算法等内容与当前的技术发展与应用需求相吻合，具有一定的理论意义和应用价值，有望在未来十年内得到进一步的应用和推广。