Convolutional neural codes for image retrieval
Xinyu Ou, Hefei Ling, Lingyu Yan, Maolin Liu · 2014
Large scale image search has recently attracted considerable attention due to the explosive increase of online images. It has been shown that convolutional neural network(ConvNet) provides a high-level feature of the visual content of the image. It has recently advanced the state-of-the-art in image classification dramatically AlexNet and has consequently attracted a lot of interest within the computer vision community. Although the classification task is quite different from the retrieval task we consider, there is an obvious hope to improve the performance of deep feature by adapting them to the task, and such adaptation is the subject of the image retrieval task. Inspired by the robust ConvNet, we focus on providing a quantitative evaluation of the image retrieval performance of the features that learned from the convolutional neural network trained for image classification. First, a simple multi-stage convolutional neural network had been constructed for parameter adjustment. When different architectures are compared, an optimal architecture for feature extraction can be used to provide activation value to be mapped to binary codes as ConvNet codes for retrieval task. Then we evaluate the performance of the compressed ConvNet codes and show that a simple PCA compression provides short codes (e.g. 16bits ConvNet codes) that give best accuracy on MNIST benchmark dataset. Experimental results demonstrate that the proposed approach is superior to some state-of-the-art methods.