Research On Local Invariant Feature Based Image Classification

As digital technology has been developing fast in recent years, how to effectively organize, search and classify vast quantities of image has become a valuable research subject with image classification as one of the most important parts. In this dissertation, we discuss image classification techniques based on local invariant feature. More specifically, they are based on bag-of-words (BoW) model and probabilistic Latent Semantic Analysis (pLSA). The main work includes the followings:The proposal of a new method to build class-specific codebook based on features’ significance and a multiclass classifier based on class-specific codebooks. BoW model, which has been widely used in image classification, was proposed for the efficient use of local invariant features. Codebook is an important part of BoW, but the k-means like clustering method used to build codebook may lower codebook’s discriminative ability. The technique we introduced here can alleviate the loss of discriminative ability. The multiclass classifier based on class-specific codebooks can take advantage of the discriminative ability of class-specific codebook and lower the dimension of BoW vector.The proposal of the pLSA which incorporated spatial information and was applied to scene classification tasks. In BoW, the code words come from feature clustering, which may generate some "polysemy" and "synonymy". pLSA has the ability to solve the problem of "polysemy" and "synonymy" and has been successfully used in scene classification as an intermediate representation of images. However, it didn’t utilize the spatial information of an image which is important for scene classification tasks. To improve the accuracy of classification, we proposed a new method which incorporates spatial information coming from neighbor words and topics’ position into pLSA. Finally, an image can be represented by the position distribution of each latent topic, and subsequently, we train a classifier on the topics’ position distribution vector for each image.