基于字典学习的稠密光场重建算法

相机阵列是获取空间中目标光场信息的重要手段,采用大规模密集相机阵列获取高角度分辨率光场的方法增加了采样难度和设备成本,同时产生的大量数据的同步和传输需求也限制了光场采样规模.为了实现稀疏光场采样的稠密重建,本文基于稀疏光场数据,分析同一场景多视角图像的空间、角度信息的关联性和冗余性,建立有效的光场字典学习和稀疏编码数学模型,并根据稀疏编码元素间的约束关系,建立虚拟角度图像稀疏编码恢复模型,提出变换域稀疏编码恢复方法,并结合多场景稠密重建实验,验证提出方法的有效性.实验结果表明,本文方法能够对场景中的遮挡、阴影以及复杂的光影变化信息进行高质量恢复,可以用于复杂场景的稀疏光场稠密重建.本研究实现了线性采集稀疏光场的稠密重建,未来将针对非线性采集稀疏光场的稠密重建进行研究,以推进光场成像在实际工程中的应用.

Dense light field reconstruction algorithm based on dictionary learning

The camera array is an important tool to obtain the light field of target in space.The method of obtaining high angular resolution light field by a large-scaled dense camera array increases the difficulty of sampling and the equipment cost.At the same time,the demand for synchronization and transmission of a large number of data also limits the sampling rate of light field.In order to complete the dense reconstruction of sparse sampling of light field,we analyze the correlation and redundancy of multi-view images in the same scene based on sparse light field data,then establish an effective mathematical model of light field dictionary learning and sparse coding.The trained light field atoms can sparsely express the local spatial-angular consistency of light field,and the four-dimensional(4D)light field patches can be reconstructed from a two-dimensional(2D)local image patch centered around each pixel in the sensor.The global and local constraints of the four-dimensional light field are mapped into the low-dimensional space by the dictionary.These constraints are shown as the sparsity of each vector in the sparse representation domain,the constraints between the positions of non-zero elements and their values.According to the constraints among sparse encoding elements,we establish the sparse encoding recovering model of virtual angular image,and propose the sparse encoding recovering method in the transform domain.The atoms of light field in dictionary are screened and the patches of light field are represented linearly by the sparse representation matrix of the virtual angular image.In the end,the virtual angular images are constructed by image fusion after sparse inverse transform.According to multi-scene dense reconstruction experiments,the effectiveness of the proposed method is verified.The experimental results show that the proposed method can recover the occlusion,shadow and complex illumination in satisfying quality.That is to say,it can be used for dense reconstruction of sparse light field in complex scene.In our study,the dense reconstruction of linear sparse light field is achieved.In the future,the dense reconstruction of nonlinear sparse light field will be studied to promote the practical application of light field imaging.