Research Of Image Classification Method Based On SPM Model

In recent years, with a substantial increase in personal multimedia devices such as mobile phones and computers, the number of images and videos people post on the network has grown explosively. However, how to effectively manage and classify video and image data is still an open issue. Currently, BOW (Bag Of Words) model and SPM (Spatial Pyramid Matching) model are widely-used image classification methods, and the latter modal partitions images into different layers on the basis of the former by introducing special arrangement position information when constructing features. Therefore, the obtained feature representation will be more complete with higher resolution. In this dissertation, based on the SPM classification model, we conduct research on feature extraction, image coding, construction of visual word code-book and feature expression of image. The main contributions of the dissertation are summarized as follows:(1) We propose a constructing method of visual word code-book based on different levels of pyramid. That is to build visual word dictionary for each sub-region after image divided, so that we can describe feature information of the image at different dividing level and in different fine grit, which can effectively solve problems like large differences among image backgrounds and position change of significant classification feature.(2) Considering that extracting the single feature of image may easily lead to information loss of the original images, this paper proposes an idea that LLC encoding the SIFT and OverFeat features of the original image and then combine them together, experiment proved that the effects feature for image classification is better than using a single feature.(3) Propose an effective weighted method for image features of different levels of the pyramid layers, and the method is based on the extent of different sub-regions feature influencing image recognition under different division levels.