Research On Image Classification Methods Based On Latent Semantic Information Of Local Regions

With the rapid development of digital imaging technology, digital image grows with an explosive speed. How to organize and retrieve images rapidly and efficiently becomes a hot issue. Currently image classification technology oriented to semantic extraction faces many challenges. How to generate a more effective image classification method is still an open problem. This thesis made a research about image classification under the needs of senmantic extraction. The main work of the thesis is as follows:(1) This thesis summarizes the current research situation of image feature extraction methods and image classification methods and analyses current challenges that image classification approaches face. Besides, it also discusses image classification schemes based on probabilistic Latent Semantic Analysis (pLSA).(2) As traditional pLSA model is unable to obtain spatial latent semantic information of image and does not consider discriminative information between latent semantic topics, so we design an image classification method based on multi-scale spatial discriminative pLSA. Firstly, it decompose image in multiple scales using spatial pyramid approach and combines pLSA model to extract the latent semantic of each local block. Then, we exploit the proposed weight learning method to learn the discriminative information between different image topics. Finally, support vector machine (SVM) classifier is exploited to perform image classification. Experimental results show the importance of spatial information and discriminative information in image classification and verify the effectiveness and robustness of our method.(3) Traditional latent semantic analysis method is incapaple of obtaining spatial semantic information and the generated co-occurrence matrix has a large quantization error. As a result, we design a novel multi-scale spatial latent semantic analysis approach based on sparse coding. After exploiting sparse coding method to make features statistics on local areas obtained by multi-scale image division, we use pLSA model to excavate the latent semantic in local areas. Then we combine the latent senmantic information in all the local areas with SVM classfier to perform image classification.The experiments demonstrate that none of these three components:spatial pyramid, sparse coding based co-occurrence matrix, and pLSA, can be excluded in our method, and they jointly improve image classification performance.