Research On Attention Selection Model And Its Application

At any time human vision access a wealth of information, and do rapid and effective treatment to them. In the whole process the key factor is the human visual attention mechanisms. Human vision can do the selection and extraction of important information and ignore subordinate information from a large of information, which is called attention selection mechanism. Studying attention selection mechanism is to imitate the human visual system more effectively in computer system, in order to achieve selectivity, efficiency, intelligence and so on in computer visual part. By simulating attention selection mechanism of biological visual and depth analysis of existing models and algorithms, the thesis proposes more reasonable method for visual attention, and shows one of its application examples on the surface inspection of steels which can been summarized as follows:(1) Based on HSI color space and modulation of amplitude spectrum saliency measurement model of static image. Consider to the problem of time-consuming calculation that exist in the models based on feature and self-information and rationality of color channels division, The thesis proposes HSIPFT saliency model combined with HSI color space and modulation of amplitude spectrum, which shows higher accuracy, stronger robustness.(2) Based on modulation of 3D amplitude spectrum dynamic saliency detection model facing with dynamic background. The existing algorithms can not reflect the saliency of movement pattern from moving objects (such as:moving direction, velocity, etc.), The thesis makes use of modulation of 3D amplitude spectrum to solve the problem of the saliency of motion pattern, and proposes 3DPFT dynamic saliency detection model, which agree with dynamic attention mechanism of human visual.(3) Attention selection mechanism is applied in surface inspection on steel plate. Accounting for real time property in practical use, the thesis obtains defective regions of steel plate by using HSIPFT saliency model. Detection of defective regions shows accurate detection and high identification rate along with the introduction of attention selection mechanism.The thesis also gives the simulation experiment and the experiment results of the presented algorithms, models. The results demonstrate these methods can solve the saliency measurement of static target and movement patterns better, and make certain applications in the actual production. Compared with other algorithms, the presented methods and models get higher accuracy, stronger robustness and better similarity of biological structure.