Efficient representation for image recognition
Date of Issue2015
School of Computer Engineering
Centre for Computational Intelligence
Image representation plays the key role in image recognition. An image is usually represented by a feature vector generated using methods like bag-of-visual-words (BOV) model, Fisher Vector, sparse coding, etc. In recent years, high dimensional feature vectors have become popular. However, they usually cause severe memory problem. In this dissertation, in order to achieve an efficient and accurate image representation for image recognition, we consider three key factors: feature generation speed, discriminance, and storage. We then propose three novel computational techniques to address the impact of the three factors on image recognition: exclusive vector quantization (ExVQ), hyper-spatial matching (HSM), and feature selection for dimensional reduction, respectively. First, considering vector quantization (VQ) is the most time consuming step in many widely used feature generation methods, we propose a fast VQ method called ExVQ. In ExVQ, an exclusive nearest neighbor search is developed to quantize visual descriptors with the most similar visual codewords. While it attains classification accuracy similar to the state-of-the-art approx NN search, ExVQ is significantly faster, leading to improved feature generation speed. Next, in order to improve the feature discriminance, HSM is proposed to characterize the spatial relationship among regions in images. Spatial information is important for many feature generation methods. Generally, to compute image spatial information, images are split into several regions in the same way. Unlike existing methods that usually compute the similarity of spatially corresponding regions, HSM computes the similarity of all region pairs in two images. Thus, HSM can encode more comprehensive spatial information. With HSM, two fast SVM classifier learning strategies are proposed, which are hundreds of times faster than general purpose HSM SVM solvers. Finally, we propose to use feature selection to reduce the storage for high dimensional feature vectors in large-scale image recognition problems. A supervised mutual information based feature selection is used to compute the feature importance for feature vectors. According to the sorted importance values, feature vectors can easily achieve different compression ratios. Extensive experiments show that the proposed feature selection achieves better classification accuracy than that of the existing feature compression methods, when they use the same memory. The three proposed methods can contribute to an efficient and accurate image recognition system. ExVQ can be used with most of existing feature generation methods to improve their generation speed without decreasing their recognition discriminance. Based on the extracted features, HSM can further improve their discriminance by considering more comprehensive spatial information. Since HSM can cause possible memory problems, with the aid of feature selection to decrease the storage of high dimensional feature vectors, HSM can be safely applied on large scale image recognition problems.
DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision