GenLR-Net: Deep Framework for Very Low Resolution Face and Object Recognition With Generalization to Unseen Categories.

Matching very low resolution images of faces and objects with high resolution images in the database has important applications in surveillance scenarios, street-to-shop matching for general objects, etc. Matching across huge resolution difference along with variations in illumination, view-point, etc. makes the problem quite challenging. The problem becomes even more difficult if the testing objects have not been seen during training. In this work, we propose a novel deep convolutional neural network architecture to address these problems. We systematically introduce different kinds of constraints at different stages of the architecture so that the approach can recognize low resolution images as well as generalize well to images of unseen categories. The reason behind each additional step along with its effect on the overall performance is thoroughly analyzed. Extensive experiments are conducted on two face and object datasets which justifies the effectiveness of the proposed approach for handling these real-life challenging scenarios.