Cross-resolution deep features based image search

Abstract

Deep Learning models proved to be able to generate highly discriminative image descriptors, named deep features, suitable for similarity search tasks such as Person Re-Identification and Image Retrieval. Typically, these models are trained by employing high-resolution datasets, therefore reducing the reliability of the produced representations when low-resolution images are involved. The similarity search task becomes even more challenging in the cross-resolution scenarios, i.e., when a low-resolution query image has to be matched against a database containing descriptors generated from images at different, and usually high, resolutions. To solve this issue, we proposed a deep learning-based approach by which we empowered a ResNet-like architecture to generate resolution-robust deep features. Once trained, our models were able to generate image descriptors less brittle to resolution variations, thus being useful to fulfill a similarity search task in cross-resolution scenarios. To asses their performance, we used synthetic as well as natural low-resolution images. An immediate advantage of our approach is that there is no need for Super-Resolution techniques, thus avoiding the need to synthesize queries at higher resolutions.