基于分块动态压缩感知的视觉处理方法研究

This project will study the blocking online dictionary learning method to solve the bottle neck of image restoration and object recognition. The sparse restrict are handled by convex optimization, greedy algorithm, and other combinatorial methods. The project is about the research of visual perception based on sparse representation and optimization methods. Firstly, the redundant dictionary is learned from blocking online dictionary learning method based on signal sparse representation model. This redundant dictionary will lay the foundation for following visual perception applications. Secondly, the image restoration is researched combining the local and nonlocal model using the learned redundant dictionary. Thirdly, the representation coefficients of different type images are fused and then the fused image is restored from redundant dictionary. Lastly, for high dimensional image data, the more stable preserve projections are studied based on dissimilarity. The projection can be applied to the real-time online target detection, recognition and tracking. In this project, effective theoretical and algorithm support are provided for studying visual perception issues. New ideas are also offered for unmanned vehicles, robots and other visual navigation systems.

针对目前图像恢复质量和目标识别准确率的瓶颈问题，本项目拟提出分块动态压缩感知方法，利用凸优化和贪婪算法等组合方法对稀疏约束进行优化求解，研究基于动态稀疏表示和优化理论的视觉感知方法。首先，项目根据信号稀疏表示模型研究基于分块在线字典学习（BODL）方法，学习到的冗余字典为视觉感知应用奠定基础；然后，利用前面学习得到的字典，结合局部和非局部模型对视觉场景进行恢复；其次，对不同模式的待融合图像进行先对稀疏表示系数融合，后根据字典恢复，在融合的同时也有去噪的效果；最后针对高维的图像数据，研究更加稳定的基于非相似度稀疏保留投影，进而应用在实时在线的目标检测、识别和跟踪方面。本项目的研究将为视觉感知系统的研制，和在无人驾驶车辆、机器人等视觉导航应用方面提供新思路。

针对训练样本不足、图像恢复质量和目标识别准确率的瓶颈问题，本项目着重面向分块的动态压缩感知方法，利用凸优化和贪婪算法等组合方法对稀疏约束进行优化求解，研究基于动态稀疏表示和优化理论的视觉感知方法。1）项目根据信号稀疏表示模型研究基于分块在线字典学习（BODL）方法，学习到的冗余字典为视觉感知应用奠定基础；2）利用前面学习得到的字典，结合局部和非局部模型对视觉场景进行恢复；3）对不同模式的待融合图像进行先对稀疏表示系数融合，后根据字典恢复，在融合的同时也有去噪的效果；4）针对高维的图像数据，研究更加稳定的基于非相似度稀疏保留投影，进而应用在目标检测、识别和跟踪方面。本项目的研究为视觉感知系统的研制，和在无人驾驶车辆、机器人等视觉导航应用方面提供新思路。