Feature Selection Methods For Image Classification

With the rapid development of Internet, multimedia technology and computer vision, massive digital image processing becomes increasingly challenging problems, namely how to quickly find the information we need from such a complicated image. In recent years, mainly in the bag of words model (Bags of Visual Word, BOV), convolution neural network (Convolutional Neural Networks, CNNs) as the representative image classification algorithm exhibited excellent classification performance has attracted more and more researchers to join the ranks.Image classification has become the key technology of organization and management of image data, but because of inter-image as well as the diversity and complexity of the existence of class differences within the class, how to be more accurate and comprehensive representation of the image is troubled by our problems, but also inspired many scholars put forward new technology to solve, promote the development of computer vision. The current research work of image classification is mainly focused on the feature representation problem, In fact, there are thousands of local features in the feature extraction stage, in these local features, not all of the features are useful. For example, the features that are located in the image background area, it generates redundancy and interference. Therefore, how to make an effective feature selection to obtain more discriminative local features and provide useful information for the image classification is an important problem in the research of classification.In view of the above problems, we proposed two methods of feature selection based on the BOV model, one is salient region-based feature selection method, the other is object detection-based feature selection method. The proposed methods combine to two feature extraction strategies:DOG SIFT and Dense SIFT. Finally, the extracted features aggregate into Fisher vector, to achieve improvements in image classification.The main work of this paper includes the following aspects:(1) We propose salient region-based feature selection algorithm, which uses the salient region to form a rectangle surrounding the box. The key point is detected utilizing the extracted box surrounded by DOG-SIFT feature and generate Fisher vectors image classification and analysis. Similarly, we apply the salient region for Dense-SIFT Fisher feature vector to perform image classification task. The experimental results shows that our proposed method achieves the good performance.(2) We propose object detection-based feature selection algorithm for DOG-SIFT and Dense-SIFT. The experimental results also demonstrate that the proposed strategies can improve the performance of image classification.