Fundamental Research On Image Saliency Detection Moedls And Methods

With the rapid development of the information processing techniques, people want computers to process and analyze the expanding images and videos and to obtain useful information from large amounts of data fast and easily. Thus, new requirements and challenges are put forward to the theories, methods, and techniques of image processing and pattern recognition. In the image processing field, computers are needed to analize images and output results. By the means of pattern recognition, image processing aims to achieve the final task of content understanding. In order to process media data intelligently, the visual perception and attention mechanisms have drawn increasing attention of researchers. Attention is a cognitive ability of selectively concentrating on the important parts in a complex visual environment while ignoring the others. It allows the brain and visual system to break through the information-processing bottleneck because the human visual system only efficiently processes parts of the massive sensory incoming information in detail. In the fields of image processing and pattern recognition, the aim of image saliency detection is to compute a kind of saliency maps which can reflect the attentive locations. Hence, using the cognitive mechanism and artificial intelligence technologies to detect and extract salient regions in images and to develop new intelligent methods is an extremely important research area in the academic nowadays. Therefore, the fundamental research on saliency detection is significant to increase the efficiency of image processing and to improve the accuracy of pattern recognition.Image saliency detection aims to solve two problems, i.e., predicting human fixation and detecting salient regions. By analyzing the deficiencies of the existing work, this dissertation points out some problems that are needed to be solved urgently. Then, the fundamental researches on image saliency models and methods are performed in the following acpects, i.e., utilizing the attention mechnism of the human visual system, solving the area-depending issue in the global-contrast-based methods, enhancing the consistency of the saliency in object regions, and embedding the general semantic characteristics of objects into the model. The contributions are as follows.1. The local-color-contrast-based saliency detection is researched firstly. The “center-surrouding” mechanism of visual attention and the spatial frequency response of the human visual system are used to build a biologically inspired model, i.e., the central-stimuli-sensitivity-based model. Furthermore, the model searches the saliency support region in order to mimic the maximal response of the receptive field in the neurophysiological experiments. Thus, the neurophysiological results are combined organically with the computational model.2. In order to solve the issues of the existing methods, a global-contrast-based vector model and a joint spatial-color constraint saliency model for pixel level saliency detection are built, respectively. In the former, a vector model is derived from the global-contrast-based saliency detection model firstly and two principles are proposed in order to effectively detect saliency using the vector model. Furthermore, by using the spatial distribution information of images, the construction of the feature vector and mean vector are optimized to solve the area-depending issue of the existing methods. In the latter, spatial and color constraints are imposed to the traditional global model and a two-layer saliency structure is proposed to generate pixel level saliency maps. The generated maps not only are with full resolution and defined boundaries, but also can suppress background effectively and improve the detection performance.3. In order to solve the issue that the generated saliency maps are not uniform enough to highlight the whole salient objects when the image is over-segmented, a multi-scale segmentation based method is proposed to extend the pixel level saliency maps to the region level. By using multiple segmented results in different scales and saliency values of pixels, the method computes the saliency values of regions in each scale. By using the image guided filtering and other ways, the proposed method improves the consistency of region level maps obviously.4. For the research on top-down saliency detection, the dissertation proposes to adopt two detection results of the general objectness and the background with low rank characteristic as high level features for learning. Thus, the general object semantic information is embedded into the model to increase the accuracy of human fixation prediction in complex scenes effectively.5. For the application research, the saliency detection model is applied to compute the saliency value of a superpixel. Then, an unsupervised method is proposed to generate semantic superpixels by using the detected results. The superpixels are merged according to the saliency values of them. Thus, more object symantic characteristics are contained in the merged superpixels. This helps the understanding of the semantic contents in images. The saliency detection model and superpixel merging method proposed in the dissertation are applied to natural images. And the expected results are obtained to show the feasibility and effectiveness of the models and methods.