Research On Generation Mechanism Of Attention And Visual Attention Calculation Model

Attention is the first key link of the intelligent system, and in a particularly important foundation position. An intelligent system without attention capacity would be drowned in the vast information ocean of the external world. Thus the study on the generation mechanism of attention has important academic significance and practical value. Moreover, because about80%of the information from the outside world is received and processed by the vision system, the visual attention should be more worthy of study.There are a lot of researches on the attention, such as Broadbent’s filter model, Treisman’s attenuator model, and so on. However, all these models are only studied to achieve a variety of possible implementations of the "selective attention", but did not reveal the generation mechanism of the attention. For example, the filter model did not explain clearly what is the filter criteria should be exactly according to? The attenuation model did not make it clear what is the attenuate criteria should be exactly according to? And so on.This study found that:various capacity of intelligent system (of course including the attention capacity) are actually servicing for the purpose (or goal) of the intelligent system. Thus, whether the external stimulus is compliance with the system’s purpose (or goal) or not should be bound to become the fundamental criteria of the attention. If the external information is relevant to the purpose (or goal) of system, we should pay attention to it. Otherwise, the information should not need to be concerned. With this new discovery, we realized:the Comprehensive Information Theory proposed by Professor Yixin Zhong can be used to reveal the generation mechanism of the attention. The Comprehensive Information Theory said:Ontological Information is the things present state of motion and its changes way, Comprehensive Information is the epistemological information which consists of syntax information, semantic information and pragmatic information. Among of three information components, the pragmatic information describes the value or the utility of the information with the system’s purpose. That is to say, the more the value of the pragmatic information is, the bigger the utility of thing to the system is, whereas smaller. Therefore, the pragmatic information can be considered as the basis of the attention. The first innovation of this thesis is:research and elaborate the generation mechanism model of attention under the guidance of Comprehensive Information Theory, and discuss the relationship of this model with the existing models.Secondly, three visual attention calculation models are proposed in the thesis, which are illustrated as follows:(1) Through the study on the fusion method of the different features, a visual attention calculation model based on feature weighted is put forward. Different features have different contribution to the object significant, such as the color make greater contribution compared to its shape for the objects with bright colors, and the shape is even more important than color for the objects with distinct angular. Thus, the different feature channels should have different weight coefficients and the weights should be set according to the features of the objects themselves. This thesis bring forward a method which calculate the respective weights of the features according to the object’s features itself, and then adjust the integration process of all saliency maps. Experiments on Amsterdam Library of Object Images (ALOI) and its extension library obtain better recognition results.(2) Through researching the influence of the feature space distribution to the model, a new visual attention calculation model based on Gaussian mixture distribution is proposed. Although the objects in the real world are in a variety of gesture, the visual system can build a relatively stable target model for the object. The diversity of the actual sample makes a single Gaussian assumption becoming unreasonable. But some traditional visual attention calculation model is build based on the simple Gaussian distribution assumption, or even don’t make the assumption, they detect and identify the target only by visual saliency point. That is why the recognition rate of the current visual attention model is not high. Therefore, in this thesis, the improved model replacing the original single Gaussian model by Gaussian mixture model (GMM) can fit to the target better, so as to improve the recognition rate. Experiments on ALOI have obtained better results.(3) Through analyzing the importance and the representation of the color features, a visual attention calculation model based CIELab color space is proposed, and successfully applied in the task of detection and recognition of traffic signs. Among the stimulation of the human visual system received, color is the most important feature. The common color spaces include RGB, CMY, HSX and CIELab, in which the design idea of CIELab space is most consistent with the color perception mechanism of the human visual system. Unfortunately, the current visual attention calculation models mostly take the RGB color space as the feature. Therefore, a visual attention calculation model based on the CIELab is presented. In the model, the color feature space is transformed from the RGB to CIELab, and features are designed specially to make the model more sensitive to color contrast information. The Inhibition of Return is controlled under a dynamic threshold, so that the model can dynamically adapt to the color changes of the samples. The model achieved very good results in the task of identifying target with rich colors (such as traffic signs). The model has advantages of bionic, without segmentation and strong recognition efficiency.