恶劣气候环境下的图像复原技术研究

When the visual system capture the outdoor image, its ability on anti-environment interference is the one of the most important factor to obtain high quality image, which also decided the results of many high level tasks, such as image recognition and image understanding. Considering the climate in northeast, our project addresses the image restoration within rain or snow. We focus on the following several aspects. Firstly, we study the appropriate filtering method to decompose the observation to smooth and detail part. Secondly, we take the scene textures and environment interference as the morphological components and establish the corresponding blind source separation model. Motivated by the morphological component analysis, we also address the dictionary learning based on relevance coding, which can promote the discrimination between the different components. And then, the mixing matrix estimation is also studied to gain the good performance on solutions and convergence. Meanwhile, as to the weak blurring and noise in smooth part, a novel approach based on hybrid regularizations is proposed, which can develop incorporating constraints within multiple priors. At last, we can obtain the high quality latent image by compositing the texture and smooth components.

在获取户外图像时，视觉系统的抗气候环境干扰能力是高质量成像的重要影响因素之一，其成像质量也决定了图像识别与理解等高层处理任务的效果。针对东北地区的气候特点，本课题重点探究雨、雪两种气候环境下的图像复原问题。其具体研究内容拟在以下几个方面展开：首先，研究一种适当的滤波方法将观测图像分解为高频细节图像及低频平滑图像；其次，将图像中的场景纹理与环境干扰作为图像的多重形态分量源，建立针对雨雪图像细节分量的盲源分离模型，并受形态分量分析方法的启发，研究该模型下基于相关性编码的字典学习问题，提高字典在不同形态分量下的类间鉴别能力。再次，探究分离模型中的混合矩阵估计方法，提高模型中优化目标的解精度及收敛速度。针对低频图像的弱模糊现象，探究基于混合正则化的复原方法，发挥多种先验信息的联合约束作用，进一步提高复原质量。最后，将场景纹理与低频分量合成，获得雨雪气候环境下的高质量复原图像。

场景图像的环境干扰复原是底层计算机视觉任务的研究热点之一，近年来受到国内外学者的广泛关注。由于该项技术在自动驾驶、智能交通监控等现代人工智能领域的广阔前景，在理论和应用层面对高质量的环境干扰图像复原技术均有迫切的需求。因此，本项目围绕这一问题开展深入的研究，针对复原问题的关键技术提出了多种创新性的解决方案。针对前景干扰与背景的分离问题，提出先利用非下采样Contourlet变换进行高、低频分量粗分离，再由高频分量提取干扰分量的两阶段处理方法。针对低频图像存在的弱模糊及噪声问题，提出了基于低秩约束的非局部稀疏编码方法和基于联合正则化模型的去噪算法。针对高频分量的环境干扰分量提取，分别提出基于形态分量分析和稀疏表示的场景建模方法和基于高斯混合模型学习的场景建模方法。通过在合成数据集与真实数据集上对上述算法进行对比验证，展示了提出的方法相比现有先进方法的优越性。