Image Representation On Set In Image Classification

The essential problem of image classification is to identify objects or targets in the image,which requires accurate description of the visual information of the image. Local information on the background details, lighting and other external conditions has good robustness and it become the mainstream of feature representation, especially Scale Invariant Feature Transform and various improved algorithm based on Scale Invariant Feature Transform appears. However, the number of local features of different images is often different, not suitable for subsequent operation such as classification and retrieval directly on local features,so we need a unified set of representation in the local features of the image collection.The set representation use a certain method to operate all the local feature points extracted from the image, forming a vector to represent the image. The main work and contributions in this paper are as follows:First, in this paper, we introduce details of three methods of set representation, named Bag of Words,Efficient Match Kernels and Vector of Locally Aggregated Descriptors. And a large number of experiments are done based on these methods on selected database, verify the three set representation classification performance.Second,we test the applicability of the Vector of Locally Aggregated Descriptors latest proposed with different clustering algorithms and clustering center numbers. According to the various of the numbers of cluster centers selected way and assignment of local features, we select K-means, Affinity Propagation and Gaussian Mixture Model, which are three kinds of clustering algorithm.Finally, we propose an improved method of Vector of Locally Aggregated Descriptors. In this paper, we study the normalization and pooling on the effect and validity of Vector of Locally Aggregated Descriptors. We choose power-law and L2 norm in the Normalized method, and use sum pooling, average pooling and Generalized Max Pooling in the pooling method.PPMI, Caltech-101 and Scene-15 are three databases about action, objects and scenes respectively. We verify the effectiveness of the above methods on the three databases.