Application Of Image Classification Based On Fisher Vector With Feature Selection

As a hot issue in the field of computer vision, image classification has aroused widespread concern in experts and scholars. Especially, in recent years, the growth of the number of digital images is explosive. Image classification has become a critical task in many fields, and many classification techniques have been proposed.At present, many state-of-the-art image classification systems follow the popular image classification scheme. First, for each image, local descriptors are extracted. Then these features are coded into higher dimensional vectors. Next, the coded vectors are pooled together, typically using some spatial pooling techniques, to form the image-level representation. Finally, a suitable classifier is used to assign the given image to a label. But many researches are focused on the last three steps, namely on the improvement and innovation of coding, pooling techniques and classifiers designation. In this paper, our research work focuses on the feature extraction step which aims at choosing the strongest discriminative subset of local features to aggregate into fisher vector. One reason is that the configuration parameters of images acquisition devices and the level of photographer are uneven, leading to great difference between images and a lot of noise features. Another reason is that most of features from background and irrelevant image segments are uninformative, some of them may even have disturbing effect. If these features are treated as a part of the global image feature, it will have an adverse effect on subsequent work with classification accuracy reduced and computation complexity increased.For the above problems, this paper puts forward two methods of adaptive feature selection in order to improve the performance of image classification tasks. One feature selection method is based on Bayesian adaptation, and the other is based on salient regions extraction. The main work and innovations are as follows:1. We develop a manually feature selection system which is used to select features manually. And then these selected features are aggregated into a global image feature for classification experiment. The results verify the effectiveness of the idea of feature selection for improving classification accuracy.2. We propose an adaptive feature selection method based on Bayesian adaptive algorithm, where we only keep the features with strong discriminative power. Then by using fisher coding technique, we get the global image expression and experiment on the dataset of Caltech256, PASCAL VOC2007and BMW. Finally, we analyze the applicability and scalability of the algorithm.3. We provide a new algorithm based on salient regions extraction to construct ideal local feature subset, and then generate fisher vector to experiment and analyze the causes according to the results.Experiments show that the fisher vector considering two adaptive feature selection methods has more discriminative power. So combining with simple linear classifiers can achieve better performance. More importantly, two proposed feature selection methods could be easily applied to the classification tasks of low resolution, low quality image datasets.