Model and metric choice of image retrieval system based on deep learning

In the era of big data, Content-Based Image Retrieval combined with deep learning technology gradually becomes the mainstream. This method can overcome some drawbacks of traditional CBIR, but at the same time there are still some problems to be solved, such as: The extracted feature dimension (generally more than 2000) is higher, which is not beneficial for efficient data storage and fast real-time query on a large scale; And the measurement method for feature matching is difficult to be determined, since the typical method based on distance is designed in the low dimension, but in high dimensional space curse of dimensionality can make those former methods may be no longer suitable. In this paper we discuss a fast efficient CBIR to improve the performace of image retrieval system, through contrasting different model and metric choices. Contrast experiments show that 4 kinds of distances (namely Euclidean Distance, Minkowski Distance, Cosine Distance, Pearson Correlation Distance) is more suitable for processing the similarity measurement on high-dimensional feature, and 5 kinds of models ( namely vgg-m-128, vgg-m-1024, vgg-m-2048, vgg-verydeep-16, vgg-verydeep-19) have higher average query precision for image retrieval task.