Research On Theories And Methodologies For Saliency Detection From Image

Human have the ability to quickly locate the most attractive contents in the large and complex static or dynamic scenes, which is referred to saliency detection. The contents are so called saliency. Saliency usually contains features that have large contrast to the surroundings, such as some signs of dangers and hazards. With the saliency detection ability of human visual system, we can quickly focus on some parts of contents in visual scene, and ignore the other background parts, thus let us prioritize external visual stimulus, and quickly analysis and make response to the most important one.With the development of image and video capturing devices, data becomes much larger and the contents of data become more complex, which correspondingly make previous computer vision algorithms hard to incompetent to current tasks. Toward this end, people try to design algorithm to simulate saliency detection ability of human visual system to firstly locate parts of important contents for subsequent analysis and treatment and ignore the redundancy, and thus facilitate other computer vision tasks, such as object detection and recognition, image segmentation, image and video compression, image retargeting, video summary, visual tracking, image retrieval and editing. Due to its importance, saliency detection has received intensive research attention in recent years, and many saliency detection models have been proposed one after another.In the field of saliency detection, based on different tasks, two branches have developed, which are generic saliency detection and specific saliency detection. According to the type of results, each branch can be further divided into two classes, visual saliency detection and salient object detection. Generic saliency detection is to find the most attractive regions and objects that draw the human attention in the natural image, while specific saliency detection usually seeks some types of regions or objects according to specific task requirement, such as face in the photo, car in the surveillance and tumor in the medical image.In this paper, we study the theories and summarize the models of saliency detection, and propose our new saliency detection features, models and evaluation metrics. The major contributions of the paper are:We survey the development of visual saliency detection models and salient object detection models, and thus conclude that both models from two classes have three major components:feature contrast fashion, saliency extraction direction and cue integration way. We also summary the saliency features, datasets and evaluation metrics of saliency detection models, and find that these models have many things in common which further prove the close relationship of two classes.We propose a generic salient object detection model. In this work, we explicitly propose two important saliency cues (features) focusness and objectness. Focusness can be estimated by scale-space analysis, while we compute objectness through modified previous object detection method. Thereafter, we combine focusness and objectness with uniqueness for salient object detection. More importantly, based on the complementary nature of the three saliency cues, we combine them in an effective and efficient way to detect and segment salient objects, which leads to the top performance when compared with the state-of-the-arts on two widely used public benchmark image datasets MSRA1000 and BSD300, by uniform evaluation metrics.We propose a way to promote diffusion-based salient object detection models. In this work, we first summary the existing diffusion-based salient object detection models, on which we further make a new interpretation of working mechanisms. We find that the performance of these models is determined by diffusion matrix and seed vector together, but the performance ceiling is fundamentally decided by the diffusion matrix. To this end, we propose methods to re-synthesize the diffusion matrix and reconstruct the seed vector to make model more precise and efficient. Previous diffusion-based salient object detection models mainly focus on how to generate good seed vector, but through our comprehensive experiments, including visual saliency promotion and our proposed constrained optimal seed efficiency (COSE), it has proved that our re-synthesized diffusion matrix can propagate saliency information on seed vector to the whole salient object more accurately. Besides, with our re-synthesized diffusion matrix, we can promote visual saliency detection models for the task of salient object detection. Our final refined model yields best performance compared with previously reported models on the benchmark image datasets MSRA10K and ECSSD, by uniform evaluation metrics.We propose a specific salient object detection model. This model implements an algorithm to to detect and delineate the tumors in breast ultrasound images. This model first locates all the tumor region candidates by learnt AdaBoost classifier, and then use learnt SVM classifier to further screen the candidates to get the true tumor regions and false tumor regions, from which we further extract foreground and background seeds. At last, we use Random Walks segmentation algorithm with located seeds to obtain the tumor boundaries. The final model proposed by us can detect and delineate the tumors very accurately in breast ultrasound images, even is applicable when ultrasound image has multiple tumors.