Research On Scale-Robust Object Recognition

As one classic technique in the field of computer vision,object recognition aims at the recognizing of the object region with high accuracy.In engineering technology,it could be applied to the fields of national defense,education,security as well as entertainment;in academia,its essential algorithm could contribute to image retrieval and image classification as one supporting algorithm.Owing to its importance in engineering and academia,more and more scholars and technicians have done some research on object recognition,and remarkable achievements in scientific research and technology have been acquired.However,with the continuous improvement of application requirement,the requirement for object recognition is also improved continuously.In practical application scenario,there exist significant differences in the object image to be recognized because of changes in imaging conditions,for example,the object image shows significant scale differences under different distance or focal length,thus making it really difficult for traditional object recognition method to recognize the image above accurately.Aiming at this problem,deep research on object recognition algorithms based on feature matching and bag-of-words model has been conducted in this thesis,improving original algorithms and proposing innovative models.The main research content and innovation points are as follows:(1)Aiming at the problem that there is no scale information in feature points using object recognition algorithm based on feature point matching,one simple Gauss pyramid based on scale information enhancement model is proposed in this thesis according to the theory of scale space.On the foundation of it,the sampling pattern and descriptor expression of FREAK descriptor are improved,according to the mechanism of retinal processing,which makes the improved descriptor own better divided ability and describing speed.Moreover,in the step of feature matching,the process of object recognition is speeded up by the establishment of the accelerated feature searching frame based on hash-index-structure.Thesis did massive experiments on image database,and the experimental results confirm the effectiveness and robustness of our algorithm adequately.(2)Aiming at the problem of poor scale robustness in traditional bag-of-words model,an improved scale-robustness object recognition algorithm is proposed in this thesis based on bag-of-words model.At first,object recognition algorithm of the spatial pyramid matching model is achieved,in which the characteristics of scale invariance are chosen to enhance the scale information.Moreover,aiming at the difficulty of inaccurate recognition when there are obvious scale differences between recognition object and object in the database,one improved bag-of-words based scale robustness recognition algorithm is proposed,in which the object scale information is increased by changing the scale of the image,which directly contributes to making the improved object recognition algorithm more robust.At last,based on the theory model built in this thesis,the improved object recognition system on the platform of Android is achieved.At the same time,thesis did massive experiments on image database,and the experimental results confirm the effectiveness and robustness of our algorithm adequately.