Research Of Image Classification Method Based On BoW Model

With the deepening of digitized degree in modern society, visual information is widely used in people’s life in the form of digital images. As an important part of image processing and computer vision, image classification can analyze and manage digital images fast and correctly to help users to obtain useful visual information. Now, image classification methods are widely used in the establishment of the library, medical image processing and unmanned navigation and other fields. In recent years, Bag of Words (BoW) model is used in image classification for its simple and effective advantages, and it becomes the mainstream technology in the field of image classification gradually. Therefore, we study the method of BoW image classification deeply, and on this basis, the key parts of the model are improved and the accuracies of classification are raised. The main research contents are as follows:1. In view of the conventional BoW model depending on single image feature which can not describe image information completely, a feature extraction method of Laplace spectrum based on uniform partitioning is proposed. We combine image scale invariant feature transform (SIFT) and Laplace spectrum to construct a more complete visual dictionary. After the feature quantification, the image spatial structure information can be described better.2. A new method of constructing visual dictionary is proposed base on a weighted hierarchical K-means clustering strategy. It uses SIFT and Laplace spectrum features to cluster the image database hierarchically to get sub visual dictionary of each category of images; then clusters the sub visual dictionaries to get the parent visual dictionaries of two features; At last, weights are introduced to two parent visual dictionaries to balance two kinds of image features’effect in image classification process. Experimental results show that the visual dictionary constructed by this method can generalize the image information better and improve the accuracy of image classification.3. According to the idea of Spatial Pyramid and multiple support regions, we propose a new Towel Center-Symmetric Local Binary Pattern (TCS_LBP) descriptor. In the process of building TCS_LBP descriptor, we use the Maximally Stable Extreme Regions (MSER) algorithm to detect image invariant feature regions; Then the detected regions are processed by Scale and affine normalization; After that, we introduce the tower zoning to the feature space to calculate the Center-Symmetric Local Binary Pattern (CS_LBP) of each divided sub-block; Finally, we combine the calculation results of each sub-block regularly to form the new local image features TCS_LBP. Image classification experiment results show that the constructed descriptor has good discrimination and can further improve the accuracy of image classification.4. To solve the shortage of traditional Spatial Pyramid Matching (SPM) methods in feature pooling, a Multi-adaptation Spatial Pyramid Matching (Ma-SPM) strategy is proposed. The Ma-SPM strategy introduces finer multi-resolution and multi-level division method to the image space. This new strategy not only retains the ability of describing image details, but also strengthens the integrity and completeness of image feature cluster. In the BoW model, we use newly constructed TCS_LBP descriptor as the image feature, and apply the new Ma-SPM strategy to quantify the TCS_LBP features, finally, the image visual histogram representations are substituted into the classifier to complete the task of image classification. The experimental results show that this method has higher classification accuracy.