Fine-Grained Bird Classification Based on Low-Dimensional Bilinear Model

Bilinear model has been shown to be very effective in a variety of fine-grained classification tasks. For the deficiencies of high-dimensional feature, a low-dimensional bilinear model is proposed. The model saves computing time greatly as well as decreases a large number of parameters by principal component analysis (PCA), and we use singular value decomposition (SVD) to replace covariance calculation. To enhance the generality of the model in the wild, we collected millions of images from bird community and flicker websites, then a new density-distance clustering method is used to eliminate a large number of noise images. Finally, a large-scale fine-grained bird dataset is constructed, which consists of 516981 images with 1329 bird categories in China. The low-dimensional bilinear model is validated on standard small-scale CUB bird dataset and our new bird dataset. The experimental results show that the proposed low-dimensional bilinear model performs better than other competing models in term of bird recognition accuracy.