Visual Dictionary Based Study On Plsa Classifier

With the explosive increase of digital information, it is now practically impossible to rely solely on man-power for timely image data classification and management. Mature and stable digital image processing techniques are needed to deal with huge amount of image data. With the appearance of more powerful CPUs and the development of image processing techniques, image classification has been studied extensively and applied in innovative ways.The performance of image classification depends on two main factors. First, the quality of images is affected by the acquisition process. Discrepancies may exist among images of the same object due to different conditions such as light, scale, orientation and obstruction. Inappropriate filtering of noise may hide valuable information as well as add to the computational load, hence deteriorates the accuracy of image classification. Second, the selection of classifiers is greatly affected by subjective criterions. There are lots of classifiers with different algorithms and respective performances. Manually labelled samples are needed for supervised algorithms to achieve high precision. While unsupervised algorithms adaptively perform classification in their application domain. In order to achieve classification with high precision, we need to select suitable classification methods according to the characteristics of the situations in question.With the above mentioned factors in mind, the present paper studied the local features, representation models and unsupervised classifiers for images. Local features and visual dictionaries are employed in an attempt to reduce noises introduced by affine transformation, multi-scale variability, partial obstruction and subtle differences among images of the same object. Image representation models based on the visual dictionaries are also built. Based on this, probability latent semantic analysis (PLSA) has been employed to achieve automatic classification without the need of manual labelling. With the proposed classifier, the theme of the images is obtained in an unsupervised manner.A comparative study on classifiers that fall within the unsupervised category reveals that our PLSA based classifier has more stable accuracy and are more robust against noise in comparison to clustering algorithms. Experiments show that optimized parameters of image representation models need to be chosen according to the problem at hand for classifiers based on visual dictionaries. This way, high classification accuracy can be achieved in a timely fashion.