Multiclass Object Recognition Based On Features Of Interest Points

Human beings can easily recognize the changing objects in nature through their visual system. However, it is rather difficult for computers. Recently, how to help computers better build a fast and accurate discriminative ability as that of human beings is one of the research hotspots among many researchers. Therefore, in accordance with the research achievements on cognitive psychology and neuroscience, it is considered to be an attractive exploring field of building a Multiclass Object Recognition Model, based on a simulation of the system processing and information handling procedures of human beings'visual system. The HMAX Model, which is proposed by Poggio, et. al, is a ventral stream of visual cortex, and it has been improved a lot and proves to be great significance to object recognition and classification ability.In this thesis, based on recent study achievements on cognitive psychology and neuroscience, the author mainly focuses on analyzing HMAX's hierarchical visual perception mechanisms and its calculating model, with a combination of studying the function and mechanism characters of the visual cortex. This study aims at improving the availability of characterized templates, as well as raising its calculating speed. Moreover, the author also implements classification and recognition purposes to eight kinds of images targets. The major findings are as follows.(1) The author firstly analyzes the physiological features and functions of the ventral stream of different visual cortexes, HMAX's hierarchical visual perception mechanisms and its calculating model, and it is found that there are many problems such as an overweight calculating task, slow recognition speed, as well as low availability of the characterized templates, etc.(2) Due to physiological features of neurons in V4 and IT, whose visual scale is limited within certain length and range, the author defines the consistent length and range of S2 cortex of HMAX Model, so as to raise the model's recognition speed.(3) Because of the selective attention mechanism of visual neurons, the small templates in the ROI can be selected as characterized ones, in order to avoid the problems of low availability which is caused by randomly selection. The selection of interest points is decided by the number of interest points, for the purpose of coordinating the thresholds standards among different classification targets.(4) Applying a modified model to classify and recognize eight kinds of image targets in Caltech-101 database, and then making a comparison with HMAX Model about the discrimination rate and time of C2 character in each image, etc. Experiment results and analysis show that the modified model is much better than HMAX Model in terms of recognition speed and discrimination rate.