CNN Features vs Classical Features for Largescale Cultural Image Retrieval

Modern applications for cultural content enrichment and management require low delay image retrieval methods in large databases. Classical image retrieval methods are suitable for certain applications but are also known to lack the ability of generalization and their use for low delay applications for retrieval tasks is not investigated for the context of cultural heritage. As a potential improvement, we propose a new approach for large scale image retrieval that uses a pre-trained CNN as a global features extractor and a clustering model trained on these features to regroup similarly looking images. For retrieval, this model quickly identifies the closest cluster and then, the matching is only carried out for the images of the selected cluster. As a result, our approach does not require indexation and the preliminary results show that it is suitable for real-time and low delay applications as in the matching step, no heavy processing is required. Some of the suitable applications include quick image search and digital rights management.