A multi-level descriptor using ultra-deep feature for image retrieval

CNN(Convolution Neural Network)-based descriptor generation is extensively studied recently for image retrieval. CNN deep feature trained for image classification is proved to have good transferability for image retrieval task. However, building a highly discriminative descriptor with CNN feature is still an important issue. The feature of the fully-connected layer is usually used and the shallow features of an image are ignored. In this paper, we proposed a simple and effective multi-level descriptor. Firstly, we proposed a multi-level feature fusion (MFF) method to capture low-level color/texture and high-level semantic information simultaneously. MFF replaces the commonly-used "object-level" with "part-level", and the filters of convolution layer are seen as part detectors, instead of using an object detector method explicitly. The complementary nature of low-level and high-level feature benefits MFF greatly. Secondly, we trained a neural net with class information to further improve the discriminative power of MFF. Our MFF achieves good performance on public image retrieval datasets. Finally, a compressed version is proposed and achieves close performance to the uncompressed version.