Disruption and Fusion Learning for Fine-grained Image Recognition

In this paper, we propose a novel "Disruption and Fusion Learning" (DFL) method to improve the accuracy of fine-grained image recognition and acquire "sharp eyes" by training the classification model. As the first step of our method, a shuffled image is acquired by controllable shuffling, that is, replacing the block matrix of the original image in a controllable range. This step destructs global structure information in the image but retain local details, forcing the model network to focus on the distinguishing local areas to get useful local feature information for recognition. At the same time, global feature information is extracted by the original image with the same structure network. Those two steam of feature later are computed by knowledge distillation part, which distills and concentrates the features both from the original image and the shuffled image to obtain effective features that help improve the recognition rate of the model. By conducting experiments on three widely used fine-grained image recognition datasets (CUB-200-2011, Stanford Cars, Stanford Dogs), our method achieves state-of-the-art performance.