A Feature Learning Approach for Image Retrieval

Abstract

Extraction of effective image features is the key to the content-based image retrieval task. Recently, deep convolutional neural networks have been widely used in learning image features and have achieved top results. Based on CNNs, metric learning methods like contrastive loss and triplet loss have been proved effective in learning discriminative image features. In this paper, we propose a new supervised signal to train convolutional neural networks. This step could ensure that the features obtained are well differentiated in space, which is very suitable for image retrieval task. We give an example on MNIST to illustrate the intent of this loss function. Also, we evaluate our method on two datasets including CUB-200-2011, CARS196. The experimental results show that the retrieval effect is fairly good on this two datasets. Besides, our loss function is much easier to implement and train.