LAUSR

Abstract Small-sample classification is a challenging problem in computer vision. In this work, we show how to efficiently and effectively utilize semantic information of the annotations to improve the performance of small-sample classification. First, we propose an image-text dual neural network to improve the classification performance on small-sample datasets. The proposed model consists of two sub-models, an image classification model and a text classification model. After training the sub-models separately, we design a novel method to fuse the two sub-models rather than simply combine their results. Our image-text dual neural network aims to utilize the text information to overcome the training problem of deep models on small-sample datasets. Then, we propose to incorporate a decision strategy into the image-text dual neural network to further improve the performance of our original model on few-shot datasets. To demonstrate the effectiveness of the proposed models, we conduct experiments on the LabelMe and UIUC-Sports datasets. Experimental results show that our method is superior to other models.