基于原型学习和线性表达的水下光视觉信息处理与理解

Underwater vision enables underwater vehicles the good ability of environmental perception in short range for further decision making, which is the key factor to its safety and intelligent level. However, due to the complex environment of underwater optical imaging, the captured images have seriously degraded, which poses great challenges for precisely sensing underwater environment and the targets. In view of the characteristics of underwater imaging, the project discusses how to effectively improve the capability of underwater visual information processing and understanding based on prototype learning and linear representation. The research content covers: with the view to image blur, color distortion and uneven illumination distribution, we propose a color correction method based on illumination-coherent regularized regression, which is further integrated into the image enhancement framework; to obtain more image details, we propose a super-resolution algorithm that combines dictionary prototype neighborhood and linear regression; to precisely identify various underwater scene images and improve the semantic distance measure between samples, we propose a metric learning algorithm using regularized linear discriminant analysis and non-negative linear regression; in view of the specific underwater targets, we propose to devise the appropriate regularized terms incorporating prior knowledge and construct the prototype dictionary to improve the performance of linear representation based classifiers. The research of this project can provide theoretical and experimental basis for the development and application of underwater visual techniques, and can be used in the field of marine military, marine resource development, and marine environment monitoring.

水下光视觉有助于水下航行器近距离内高效感知环境以辅助决策，是保证其安全性和智能化水平的关键因素。然而，受水下复杂成像环境影响，使采集到的图像严重降质，对准确理解水下环境和目标带来极大挑战。本项目根据水下成像特点，以原型学习和线性表达为技术主线，探索能有效提高水下光视觉信息处理与理解能力的方法。研究内容包括：针对水下图像模糊、颜色失真、光照分布不均等缺点，提出基于光照一致正则化回归的颜色校正算法，进而融入增强框架；为获得更多图像细节，提出结合字典原型邻域和线性回归的超分辨率方法；为准确区分水下不同场景图像，增强样本间的语义距离度量，提出基于正则化线性判别分析和非负线性回归的度量学习算法；针对特定的水下目标，提出结合先验设计合理的正则项，构造原型字典，以提高线性表达分类器的识别精度。本项目的研究可为水下光视觉技术的发展和应用提供理论和实验依据，可用于海洋军事、海洋资源开发和海洋环境监测等领域。

针对水下复杂成像环境，本项目从水下光视觉信息处理与理解系统设计的角度出发，研究了原型学习和线性表达在水下目标识别/场景理解中的应用。主要研究成果包括：1) 在水下光视觉信息处理方面，针对水下图像模糊、颜色失真、对比度低等缺点，提出了一种基于颜色校正和非局部先验的水下图像复原方法。结合水下成像特点，该方法通过颜色校正、透射率图估计和修正，以及背景光估计等步骤，有效去除了水下图像颜色失真，大幅提升了图像对比度和清晰度。2) 针对水下图像颜色失真(呈现蓝绿色调)问题，提出了两种基于学习的颜色校正方法：基于边缘矩和光照一致正则化回归的色彩恒常算法和基于颜色边缘矩和锚定邻域回归的色彩恒常算法。这两种方法均获得了较高的光照估计精度，有效解决了图像颜色失真问题。3) 在水下光视觉信息理解方面，鉴于仅用单一特征难以准确表述水下复杂场景下目标物体的特征属性，提出了基于融合的多类特征提取方法。采用包含颜色、纹理和形状的融合特征将水下目标图像表示为一个高维特征向量，在鱼类数据集上的实验表明该方法大幅提高了水下鱼类图像识别准确率。4) 为准确表达高维特征间的语义距离，提出了两种适用于多特征融合的距离度量学习算法：基于锚定核的度量学习算法和基于参考样本对约束的核化度量学习。以上方法生成了更有鉴别力的度量矩阵，进而提高了水下目标的准确识别率。本项目的研究为实现水下图像处理能力强、水下环境识别精度高、鲁棒性好的智能化水下光视觉系统提供了丰富的理论和实验依据，可用于海洋资源开发和海洋环境监测等领域。