Aggregating Local Features of Convolutional Neural Network for Material Image Retrieval

Abstract

Large-scale microscopic images in materials science need to be indexed and managed using practical management tools. Content-based Image Retrieval (CBIR), which indexes and searches images based on the image features, allows for long-term data management in large-scale image datasets. Considering the difference between material microscopy images and natural ones, we propose a novel CBIR method for material microscopic images. In the proposed method, convolutional neural networks (CNN) are used to extract local features from an image, and the scale-invariant feature transform (SIFT) model is used to generate a keypoint density map (KDM). Experiments on a material microscopic image dataset show that the proposed method achieves an approving retrieval performance.