Font Size: a A A

Research On The Object Recognition Technology Based On Visual Cognition

Posted on:2016-04-13Degree:DoctorType:Dissertation
Country:ChinaCandidate:Y X ChangFull Text:PDF
GTID:1108330473956122Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
In machine learning, pattern recognition and computer vision and other relevant fields, object recognition has been one of the most challenging problems. Object recognition technology research, which is the basis for complex visual task such as solving the object tracking, behavioral understanding and scene analysis, has important theoretical and significance practical value, as well as become a very important focus and difficulty research hotspot Object recognition is the technology which uses the computer to process, analyze and understand images and then gain the relevant information of the research objects or region. The ultimate goal of object recognition research is to allow a computer like the visual cortex of the human brain "read" the contents of the pictures quickly and efficiently, which will lead us into the more intelligent future. It also promotes interest of researchers in studying the biological mechanisms and intelligence of visual cognition. Especially in the complex and harsh environments, traditional computer vision methods based on statistical learning encounter greater difficulties in the processing of visual information. In view of above factors, how to study and design computer vision algorithms from the visual cognition perspective becomes an urgent and challenging task.The key step of the biological visual perception system-hierarchical maximization and learning mechanism are mainly considered and focused on for research work in this paper, which is used to evaluate and analyze the optimization algorithm model performance on object classification, recognition systems, and on a static 2D image data sets and continuous frame 2D image data sets. Particularly, it studies the classical computational model based on biological vision (HTM, HAMX) and classic visual computing model. The main research contents are as follows:(1)Because traditional target recognition models have poor robustness and other problems.firstly,this paper analyze and research the HMAX and HTM two artificial brain models from the theoretical calculation and algorithm design level, and points out that the principle of correspondence between the essence and computer vision model. Then, the application of visual cognition theory is detaily analysed in the target classification and recognition.(2)Considering that the light abrupt changing leads to the problem of low detection rate, combining homomorphic filtering algorithm and more distinguishing LBP feature based patch is proposed to apply in target recognition system. The experimental results verify the effectiveness and feasibility of the optimization algorithm.(3) In order to solve the problems that detecting objects and accurately give out its visible parts is a major challenge for object detection. In this paper an explicit occlusion modeling is presented through combining appearance and motion. This modeling combines together two parts:part-level object detection with single frame; object occlusion estimation with continuous frames. It breaks through the performance bottleneck caused by lack of information and effectively improves object detection rate under occlusion. It improves the performance of partial object detectors through rescoring with semantic parts.(4) Owing to the problems that the traditional feed forward hierarchical model of the visual cortex is limited to solve the classification problems, this paper proposes a novel feed forward hierarchical model for an objective detection from multi views. The model adopts the objective expression method based on perspective. By adding a simple unit layer to express different perspective of target and adding a complex unit layer to express the voting results of different simple unit perspective, the model achieves the target detection of arbitrary poses. In the aspect of learning, based on the patch learning of bottom layer feature, The model adds on learning perspective at the high layer, therefore forming a two layer learning structure. The structure can effectively improve the efficiency of learning. The test results on PASCAL VOC 2011 datasets demonstrate that compared with the performance of traditional computing model; our model achieved better detection effect.(5) In order to solve the challenging problems of recognizing object in the angles changing and occlusion, a novel multi-angle algorithm is proposed by combining the Gabor feature and shared LIOP feature during the changing poses. First of all, the input image is filtered by a 2D Gabor filter in 4 directions and 16 scales to obtain 64 groups of characteristic response map. Then the scale and translation invariant feature can be derived from computing the maximum response value among the adjacent scales and position. Secondly, the geometric transformation algorithm is utilized to gain the shared LIOP feature under different perspectives. Thirdly, for reducing the time complexity, the dimension of combined features is reduced by the principal component analysis. At last, the calculated feature is trained and learned in SVM for the detecting model.In conclusion, aiming at the limitation problem of traditional target recognition in the complex scenes, this paper mainly focuses on using and simulating the biological characteristics of visual perception to explore and guide the key problems of the traditional target classification, recognition field and algorithm models. We have achieved the target recognition function based on visual cognition, finished the given research task, achieved initial results. Our research provides a theoretical basis for further algorithm research, also has certain reference significance to the study of biological vision and computer vision.
Keywords/Search Tags:visual perception, object recognition, complex scenes, hierarchical maximization, machine learning, multi-angle
PDF Full Text Request
Related items