Visual Attention Modeling And Its Application In Image Analysis

Attention is an important psychological adjustment mechanism during the courseof human information processing, which can process and contribute limited informationresources that enables perception the ability of selection. When human vision system isfacing complex scenes, it can quickly focus on small numbers of salient visual objects,and such kind of process is called visual attention selection. Researching on mechanismmodel of visual attention is not only helpful in understanding the working mechanism ofhuman visual system, but also has important application in image analysis andunderstanding. If the mechanism of selective attention can be introduced to imageanalysis, allocating the computing resources on the regions that are more attractive tohuman attention, the efficiency of modern image analysis methods will be greatlyimproved.This article is grounded on how to design a practical visual significant regiondetection algorithm with the introduction of the selective attention mechanism; how toreconstruct images under the guidance of the focus fixation; how to introduce attentionmechanism to the recognition and visual search of specific image (face). Generally, themajor work in this study mainly includes:1. Based on characteristics of the scale choice of human visual system, we proposea method that can calculate saliency through multi-scale analysis to the frequencydomain characteristics of images. Physiological experiments proved that people hasdifferent sensitivity threshold toward different frequency visual information, and thissensitivity threshold can be presented by Contrast Sensitive Function, CSF. This articlewill firstly conduct the statistical analysis of human fixation in image scale space,achieving a more accurate sensitivity curve and drawing a conclusion that human visualsystem is more sensitive to mid-low signals. Combining this curve, different weights aregiven to frequency band using Gaussian band-pass filter in this article, which is in orderto calculate the saliency area of images. Experiments in images and videos prove thesuperiority of this method and it has better results especially to the saliency detection incomplex repetitive texture scene. The time complexity of this algorithm is low, fastimage and video processing, and better effect in detection.2. Currently, most popular face representations are based on the uniform gridsampling; however, what is able to provide discriminant information is only a smallregion of human face. Psychologists’ studies have shown that the human visual systemgive different weights to different regions of human face via space-variant sampling onfovea and non-uniform distribution of fixation. This article reconstructs the image ofhuman face by fixation and foveated imaging. And adopt different Gaussian smoothingoperator to smooth toward different areas of images. Through analysis the effect of human face reconstruction brought by human fixation, attention focus produced by theartificial calculation model, randomly generated points and the uniform grid sampling, itverifies the effectiveness of non-uniform distribution fixation to the reconstruction offace region, which can provide strong evidence for the subsequent face recognitionalgorithm.3. Psychology experiments show that for an input image, before human fixates atthe local area, human visual system first will have a global understanding, and then thefixations will fall one after another to the regions of more interested (local). This paperintends to imitate the way of human visual system processing information, to design aface recognition method based on selective attention. By extracting the global featuresof the image to simulate the global features extracted by the human visual system beforethe generation of fixations, and then using the real fixations of human eye, we get thesaliency map of face images. Then according to the saliency map, divide the faceimages to fixated and non-fixated regions to mimic the human foveated imaging inorder to extract local features. Finally, we integrate the local and global features tofinish recognition. In this paper, the previously mentioned experimental framework hasbeen tested on public face database and general object database, the experimental resultsof which has showed that, by introducing selective attention of human visual systeminto the face recognition, not only the amount of coding has been reduced, but also therecognition performance has been improved, which proves the reasonability ofintroducing the selective attention of human visual system into face recognition.4. Changes of fixation play an important role in obtaining human visualinformation and target searching. This paper proposes a face target and itscharacteristics search system based on neuron receptive field invariant feature extractionand eye movement control mechanisms. As an example, the system is applied to thesearch of the human eye center. The main mechanism of this system: in the learningphase, uniformly set initial position of the viewpoint on images, which can make thesystem perceive and learn the face target and its characteristics of spatial orientation,distance and scale that need to search; through the control issued by high-level cognitiveneurons, it makes the fixation move from the initial position to face target and othercharacteristic positions; in the actual operation phase, given a random initial fixationposition on images, the system will automatically move to the target based on learningand memory experience in command and control fixation, through four or five stepsfixation move and finally achieve the position coincidence of fixation and targetcharacteristics (such as eye center). Likewise, the system can sequentially search all thecharacteristics of the previous target in the image, in order to achieve the target direction,the perception of distance and scale, search, and location.This article studies visual attention modeling and its applications in image analysis,research, and gives effective solutions, which proves that the reasonability of introducing the selective attention of human visual system into image analysis.