Research On Saliency Detection Via Light Field Under Complex Scenes

Computer vision is widely used in human life.However,when a computer processes a picture,it tends to process the entire image with the same algorithm,which takes a lot of time in useless regions.Saliency detection,aiming at selecting the most visually distinctive regions from a scene,has attracted much attention.A lot of computation can be saved by only performing subsequent algorithm on the region of interest.However,the existing saliency detection algorithms for 2D and 3D images cannot obtain accurate results when processing complex scenes(for example,the foreground regions have similar colors and depths with the background regions).The light field images have rich multi-modal information(including color,depth,and focus information)that can help algorithms identify salient objects in complex scenes.However,the existing light field saliency detection algorithms do not consider the correlation between multi-modalities.There are still some misdetections in complex scenes.At the same time,existing light field methods are limited to traditional algorithms because of the shortage of data.To address the above issues,this thesis proposes a traditional saliency detection algorithm and a deep learning saliency detection algorithm for light field images.Firstly,this thesis proposes a light field saliency detection algorithm based on Depth-induced Cellular Automata.The algorithm uses the multi-modal information of the light field to construct an object-guided depth map as an inducer for effectively utilizing the correlation of multi-modal information in the light field.Then,the proposed algorithm constructs an optimization model,named the Depth-induced Cellular Automata(DCA),to take advantage of the spatial consistency of an image.The saliency value of each superpixel is updated by exploiting the intrinsic relevance of its similar regions.Additionally,the proposed DCA model enables inaccurate saliency maps to achieve a high level of accuracy.The updating rule of each superpixel is determined by the object-guided depth map.This thesis then proposes a light field saliency detection algorithm based on deep learning algorithm.This thesis builds the largest light field saliency detection dataset including 1465 scenes.Then,this thesis proposes a network which contains two convolutional neural networks to extract the features of the focal stack and the all-focus image,respectively.In order to highlight the effect of the focal slice focused on the salient object in the final prediction,a recurrent attention mechanism is used to adaptively learn the feature fusion weights of each focal slice in the focal stack.Finally,the predictions of the all-focus image and the focal stack are fused to obtain the final high-quality prediction.We analyze the proposed approaches on the publicly available LFSD dataset and the proposed dataset.Extensive experiments show the proposed methods are robust to a wide range of challenging scenes and outperforms the state-of-the-art 2D/3D/4D(light-field)saliency detection approaches.