Research On Key Techniques Of Image Representation And Classifier In Image Classification

As one of the most important topics in computer vision, image classification has been playing an indispensable role in our daily life, industrial manufacture, medical facility, military activity and so on. The advent of big data equips image classification with even more potential and applications. Facing the ever-increasing image quantity and variety, great variations in such as illumination, pose, background and clutter, image classification is a very challenge problem. The common used classification framework contains three important parts, feature extraction, image representation and classifier. This work summarizes the representative methods in all three aspects, and focuses on image representation and classifier, proposing a novel representation method and classification algorithm.As for image representation, numerous works have proven the importance of preserving locality and leveraging high-order information of features when constructing an image representation. Motivated by these works, we propose to leverage high-order statistics in the local pooling (LP) method, which we call high-order local pooling (HO-LP). We first estimate the high-order statistics in both local feature space and spatial regions as Gaussian distributions. Then we study the coding method on Gaussian distributions which form a Riemannian manifold and propose a both efficient and effective method to map Gaussian distributions with diagonal covariances into the linear space, where the classical coding methods in Euclidean space can be employed to get the final image representation. The proposed image representation shows promising performance when using either hand-craft features or convolutional neural network (CNN) features.For classifier, the popular CNN feature based image representations simply employ SVM for classification while the nearest neighbor (NN) based classifiers have attracted little attention and the few attempts based on Naive-Bayes Nearest-Neighbor (NBNN) failed to show satisfying performance. In order to overcome limitation of NBNN, we propose a novel large margin nearest subspace (LMNS) classifier. Instead of matching local features as in NBNN, we introduce three image-to-class (I2C) distances in the space of image, which speed up the classifier by several hundred times. Furthermore, we propose to employ a large margin metric learning method which significantly improves the classification performance. Besides, we propose a new image representation strategy to make better use of CNN features in a very efficient way. Combining the proposed LMNS classifier and image representation, we achieve very competitive performance on various image classification benchmarks.We test the proposed methods on a variety of classification tasks. We use 8 image classification benchmarks, covering the application of scene classification, object recognition, material classification and fine-grained classification. Numerous experiments demonstrate that the proposed methods are very competitive and versatile.