Research Of Sparse Coding For Fine-Grained Visual Categorization

Fine-Grained Visual Categorization is a both interesting and important research topic. With the popularization of computer technology in people’s daily life, there is a growing need in quickly and accurately distinguishing some fine-grained categories, such as flowers or birds. However, due to the existence of large inner class differences, e.g. the deformation of petals and variation in color, and large intra-class similarities, e.g. the highly similar images from Kentucky Warbler and Magnolia Warbler. Therefore, it is not suitable to directly adopt image classification algorithms that originally designed for large scale visual targets. Hence it is of great importance for researchers to look into the details of specific fine-grained images, and to design algorithms which can accurately and automatically classify images.From the perspective of feature extraction and sparse coding, this paper mainly studies the classification methods for flower and bird species. The main contributions of this paper are summarized as follows:1. An image region alignment method for bird’s crown, beak, eye, forehead is proposed to cope with the highly diversity of bird’s poses. From observation it can be found that the parts arrangement in a bird’s head mainly remains unchanged. By calculating the position statistics of each part in a training set, a head part position in an unknown image can be predicted. And the following feature extraction procedure is performed on this aligned part region.2. Some classical feature extraction algorithms are studied including color based methods and gradient based methods, and a few common feature coding methods are also discussed. Experiments are conducted on a flower dataset to verify those algorithms’ actual classification performance.3. A local constrained two layer sparse coding architecture in an image gradient quantization space is proposed to extract middle level features. This two layer architecture aims to extract discriminative features as many as possible through continuous coding process and max pooling process. In the meantime, a local constrained dictionary updating process is derived to emphasize the smooth variance in quantifying features. Experiments show that this architecture can effectively classify both flower and bird species.According to the experimental results given in this paper, the proposed sparse coding architecture for fine-grained visual categorization can achieve high classification accuracy on flower and bird species. Moreover, for bird species, the proposed head part alignment method is very robust and effective.