Visual Salient Object Detection And Its Application

In recent years, the amount of images grow explosively with the development of mobile internet. In order to find the useful information from the huge image data col-lections, we desperately need fast and accurate image processing techniques. Due to the limitation of brain capacity, human visual system could automatically select important visual stimuli from natural scenes for priority processing. Such selective attention is an important mechanism for we humans to accommodate to the changed environment rapidly. Inspired of the selective attention, researchers propose saliency detection meth-ods to simulate the visual attention mechanism. The saliency detection methods could locate the most attention-grabbing regions in an image, which is very useful for elim-inating the interference of irrelevant image information. Saliency detection has been a popular research subject, and applied in many computer vision applications, such as object segmentation, object recognition and object tracking.Human visual system observes the environment both in a fast, data-driven bottom-up manner and in a slow, task-driven top-down manner. Most of the saliency detection works evaluate bottom-up visual saliency for its free of the involvement of high-level knowledge. In this work, we focus on bottom-up salient object detection, and propose two novel salient object detection methods based on analysing the defect of existing methods and utilizing the underlying biology principles of saliency. The main work of this dissertation could be summarized as follows:1) Based on the salient object selection mechanism of human visual system, we pro-pose a saliency biased object detection methods, in which the estimations of re-gion saliency and objectness are explicitly separated. Given an image, we first estimate the possibility of each region belonging to an object (i.e. objectness) to locate all the possible object regions. Then we compute region saliency based on contrast analysis. The salient object regions are obtained by biasing object regions with saliencies in a non-linear fusion manner. To overcome the salien-cy inconsistency of the regions belonging to the same type, we further propose a diffusion based saliency optimization process. To be specific, we select some seed regions from the initial saliency map, and learning region-pairwise similari-ties based on region-pairwise features, and then update the saliency value of each region based on its similarities with the seed regions to obtain a more consistent saliency assignment. The experimental results demonstrate the effectiveness of the proposed method.2) Based on the cause of region saliency, we propose a background driven salient object detection method. We first analyze the defect of existing local or global based methods, and discover the significant role of background in contrast based saliency estimation. Hence, we segment quasi-background regions from the back-ground map as contrast reference regions. Such background map could obtained via any background prior. We particularly propose a background learning method to predict the possibility of each region belonging to the background, in which the convolutional neural networks is utilized. For computing region contrast, the col-or and texture are both used, and their weights are determined image-dependently based on feature distributions between foreground and background regions. To improve the wholeness of the detected salient objects, an enhanced graph-based optimization process is also conducted, in which the background priors and non-local connections from feature space are embedded in the traditional k-regular graph. The saliency results are optimized by propagating saliencies along the graph based on region-pairwise similarities. Experimental results on several pub-lic datasets show the superiority of the proposed method.3) To verify the practical application value of salient object detection, we apply the proposed saliency detection methods in object segmentation and object classifi-cation. For object segmentation, we discuss the adaptive segmentation, GrabCut-based segmentation, and user interaction-based segmentation of the saliency map-s. Segmentation results demonstrate that saliency analysis could improve the per-formance of object segmentation. For object classification, to eliminate the infer-ence of image features from background regions, we segment foreground regions from the saliency map and extract local image features from these regions for clas-sification. By comparing the influences of different saliency detection methods on classification performance, we show that saliency detection could significant-ly enhance the performance of object classification, and the better is the saliency detection, the better is the classification performance.In summary, we focus on salient object detection and go deep into its study. In-spired by the relevant biological principles behind saliency, we propose two different salient object detection methods, i.e. saliency bias and diffusion based salient object detection, and background driven salient object detection, which both achieve state-of-the-art performance on the saliency and wholeness of detected objects. We also apply the proposed methods into object segmentation and object classification, which demon-strate its potential applications in computer vision.