Topic model based bird breed classification and annotation

In this paper, we propose a novel graphical model considering saliency (GMS) to classify and annotate the finegrained bird breed. The processing can be divided into four steps. Firstly, each image is over-segmented into several regions. Then, we use GMS to perform the classification and annotation based on the region level and patch level feature. To further improve the precise of classification, SVM is employed based on the features extracted from the annotated bird. Finally, the posterior probability distribution of category obtained by GMS and SVM is combined to perform the image classification. During the parameter learning phase, we use the Gibbs sampling to establish the optimized parameters of the model. Experiments on the well-known Caltech-UCSD Birds dataset demonstrate that the proposed model can achieve impressive results compared with existing methods based on topic model.