The Research On An Indoor Object Classification Method Based On Fusing Multi-feature

Object classification's essence is to classify the category of the object such as lamp, chair etc. contained in the image or video in the view of computer. In recent years, due to the high popularity of the smart mobile devices and the rapid development of the network technology, the number of the digital image has an explosive growth every day. In the face of these image information, people hope that we could use computer technology to handle these data quickly and efficiently and extract valuable information, which prompts the field of the object classification' generation.The study of this thesis is an effective fusing multi-feature method in object classification technology based on BoW model. When there are too many kinds of objects, one type of feature is not enough to express the appearances of the object, therefore, fusing multi-feature is becoming a popular research direction and has had some achievements. However, there are also some problems, for example, the time cost and the space cost will unavoidably increase, and the single feature's defect may be enlarged etc. This article's main research work is to solve the above two problems effectively as how to improve the correct rate of object recognition. Specific work consist of the following several aspects:(1) Extract the image's SIFT feature and multi-scale LBP feature. SIFT feature is an mature multi-scale feature, describing the image's gradient direction information, and could solve the weak light conditions of the indoor object very well. LBP feature describes the texture message of the image, and the multi-scale LBP feature can extract more detailed texture information. This thesis fused SIFT feature and multi-scale LBP feature, combined these two kinds of features' strengths and complemented each other. Thus the image can be described more all-round.(2) In this thesis, the concept of word bag coming from BoW was applied to LBP algorithm, and presented an approach to extract image's LBP feature based on BoW model. Compared with the frequently-used LBP feature extraction method, this approach can effectively reduce the dimension of the LBP feature vector, and then reduce the computational complexity of the follow-up object classification.(3) This thesis put forward an impactful fusing multi-feature method based on SIFT feature and multi-scale LBP feature. Compared with one kind of feature, this method could improve the accuracy of object recognition effectively. First, utilize SVM classifier to pre-classify separately on these two kinds of features and assign their weights 0 or 1 according to the size of pre-classification results. Then, the method fuses SIFT feature and multi-scale LBP feature by introducing their own weight, obtaining a fused integrated vector. At last, classify on the fused vector using SVM classifier again, and then achieve the final result of object recognition.The experimental results showed that the method proposed in this thesis could improve the accuracy of object recognition effectively.