Few-Shot Visual Classification Using Image Pairs With Binary Transformation

Accurately classifying images using few-shot samples have been widely explored by researchers. However, these methods have two drawbacks. First, images are often used independently. Second, class imbalance is ignored and hinders the classification accuracy with the increment of classes. To tackle these two drawbacks, in this paper, we propose a novel visual classification method using image pairs with binary transformation (IPBT). For one image, we bundle it with each training image into an image pair by concatenating the representations of the two images along with their similarity. The class consistency of two images is used to split the image pairs into binary groups. One group contains image pairs of the same class, while the other group consists of images pairs belonging to different classes. We train classifiers to separate the binary groups apart. To classify a testing image, we first bundle it with all the training images that are then predicted using the learned binary classifier. The image pair with the largest response is selected, and the testing image is assigned to the same class of the paired image. We conduct few-shot visual classification experiments on three public image datasets. The experimental results and analysis show the effectiveness of the proposed IPBT method.