Optimization of deep convolutional neural network for large scale image retrieval

Feature extraction and similarity measurement are two key steps in image retrieval. AlexNet is a classical deep convolutional neural network for image classification, but using it directly for large scale image retrieval is not efficient. To address this issue, we propose a novel framework to improve its ability for feature extraction and its efficiency for similarity measurement. The proposal optimizes AlexNet in three aspects: pooling layer, fully connected layer and hidden layer. In particular, average pooling is replaced by max-ave pooling for better local feature extraction; the non-linear activation function Maxout is used in fully connected layers for better global information extraction and hidden layer is added for mapping high-dimensional feature into binary codes. The proposed framework outperforms state-of-the-art methods on public databases for image retrieval, including large scale database.