A visual attentional system, playing a role in abbreviating a visual infomation, is presented. This computer model is composed of feature modules, saliency maps, and selective processing. Multi-scale images are input into feature modules and are combined on multi-scale saliency maps. Winners-Share-All network(WSA)implements the selection process based on conspicuity of locations. WSA is active every layer of saliency maps. Conesquently some locations are selected simultaneously. By using WSA and multi-scale processing, it is possible to express a spatial-temporal quality of visual attention. This model is supported to overcome a limited-capacity on human visual processing.
Assuming that 3D-objects are represented by a set of 2D-views, how the variability in object appearance caused by various factors is dealt with must be accounted for. We investigated the role of"canonical views"by psychophysical method using novel objects. In experiment 1, we showed that novel objects possessed canonical views, which were easily recognized and rated typical. In experiment 2, we examined that the effects of view canonicality on the range of generalization. The resultsshowed that the range of the generalization for canonical views was the broadest compared to other views, which supported the representation of 3D-objects based on canonical views.
We examined effects and functional roles of category information in visual cognition by psychological experiments and neural network simulation. In psychological experiments, similarities within category increased when category information is provided, and indicated that effect is a top-down category information effect. We can simulate this effect by neural network model that has bidirectional connections and made internal representations from mix of bottom-up input and top-down category bias. Top-down category bias exaggerates common features of category. Exaggerating common features is relevant to deduction from category label. This functional role of category(deducing features from category label)is thought to owe similarity changes observed in psychological experiment.
Auditory stimulus is an important source of information about surrounding objects. The Sonic-glass is one of the most excellent device that enhances the sound localization ability. However, such a passive methods inevitably provide limited sort of information, and some kind of information, velocity of the objects or concave for instance, are difficult to recognize with sound. Addition of visual information on the natural sound, by the computer synthesized sound, will overcome his difficulty. In this method, however, the compatibility of the synthesized sound with the natural stimulus is critical for inconsistent artificial stimulus may be annoying and, in worst case, even disturb the ordinary localization. In this context, localizable sound was synthesized on a PC and it's compatibility with natural sound was examined.
The present experiment investigated the processing of visual and auditory stimuli in short-term memoty. The subjects were confronted with two continuing sequences of words, the target and the test sequences, one of which was presented visually and the other aurally. They were required to memorize each of the words in the target sequence, and to recognize whether each of the words in the test sequence had been presented at some earlier point in the target sequence. The results showed that recognition performance was a decreasing function of the delay time between the presentation of the target and the test words, and was superior when the target suquence was presented with aurally and the test sequence was presented visually, relative to when the modality of the presentation of these sequences was reversed. On the basis of these results, processing of visual and auditory stimuli in short-term memory was discussed.
Focusing on the difficulties existing in the image understanding and pattern perception that includes face recognition, we select the major two key factors --- non-parametric optimization and nonsupervised mechanism --- as the basis of efforts to make any potential breakthrough in the field. In this paper we summarize a unified framework of visual pattern discovery, which has a much stronger pattern discrimination capacity than the commonly used data mining and KDD. A novel logical"tough set"concept is proposed and corresponding mathematical measures are also discussed.
Based on the emergent mechanism in the underlying perceptual processis for nonlinear patterns, we have proposed and developed a series of models for adaptive image segmentation and corresponding algorithms by emergent computation. With this approach, the degree of dependence on the prior knowledge and ad hoc environment is reduced greatly according to the optimization criterion. Our proposed Genetyllis system oriented to human-to-machine applications are also discussed. Our work is aimed at exploring any functional way to extract the key pattern features through the integration of psychological, physiological and artificial processes in terms of the emergent computational techniques in the domain of visual pattern discovery.
In this paper, we describe an approach for estimating the orientation of a signboard relative to the camera from a single view rectifying a distorted signboard image taken from a non-orthogonal viewing direction. First, signboards in scene images are detected by using several heuristics of characters and character lines. Then, we try to compute the relative orientation of the camera and a detected signboard from a single view and rectify a distorted signboard image using two methods. Experimental results indicate the effectiveness of the proposed approach.