Evolving Convolutional Neural Network and Its Application in Fine-Grained Visual Categorization

Fine-grained visual categorization is one of the challenges in computer vision due to the high intra-class but low inter-class variances. Convolutional neural networks (CNNs) are widely used to solve this problem. However, a huge number of clearly labeled images are usually required to train a CNN model for a high precision, which may be quite costly and time consuming. To overcome this problem, in this paper, a novel evolving CNN (ECNN) is proposed, which can efficiently utilize the limited clearly labeled images and a large number of weakly labeled images. The overall framework contains two parts: one for collecting the weakly labeled images from the Internet by Web crawlers; and the other for updating the CNN classifier. Specifically, several different search engines are adopted to collect the weakly labeled images, in order to get relatively comprehensive results. The proposed method is demonstrated on several datasets, including CIFAR-10, Oxford pets, and Chinese food dataset. The results show that ECNN outperforms the traditional CNN and achieves the state-of-the-art in most cases.