CNN-based methods have been proven to work well for saliency detection on RGB images owing to the outstanding feature representation abilities of CNNs. However, their performance will degrade when detecting multiple saliency regions in highly cluttered or similar backgrounds. To address these problems, in this paper we resort to light field imaging, which records the color intensity of each pixel as well as the directions of incoming light rays, and thus can improve performance for saliency detection owing to the usage of both spatial and angular patterns encoded in light field images. However, it is not trivial to use CNN-based methods for saliency detection on light field images because these methods are not specifically designed for processing light field inputs and current light field datasets are not sufficiently large to train CNNs. To overcome these issues, we first present a new Lytro Illum dataset, which contains 640 light fields and their corresponding micro-lens images, central-viewing images as well as ground-truth saliency maps. Comparing to the current light field saliency datasets~\cite{Li14, Zhang17}, the new dataset is larger, of higher quality, contains more variations and more types of light field inputs, which is suitable for training deeper networks as well as better benchmarking algorithms. Furthermore, we propose a novel end-to-end CNN-based framework for light field saliency detection as well as its several variants. We systematically study the impact of different variants and compare light field saliency with regular 2D saliency on the performance of the proposed network. We also conduct extensive experimental comparisons, which indicate that our network significantly outperforms state-of-the-art methods on the proposed dataset and has desired generalization abilities on other existing datasets.

本研究提出了一种新的基于CNN的光场显著性检测框架，提出了三个新颖的MAC（模型角度改变）块来处理光场微透镜图像，同时还建立了一个大规模的数据集和标准，可以用来训练更深的神经网络和进行评测，实验证明，我们的方法在该数据集上明显优于现有的最先进方法，并在其他数据集上具有理想的泛化能力。

基于深度卷积网络的光场显著性检测