Publication:

Aggregation of deep features for image retrieval based on object detection

Image retrieval can be tackled using deep features from pretrained Convolutional Neural Networks (CNN). The feature map from the last convolutional layer of a CNN encodes descriptive information from which a discriminative global descriptor can be obtained. However, this global descriptors combine all of the information of the image, giving equal importance to the background and the object of the query. We propose to use an object detection based on saliency models to identify relevant regions in the image and therefore obtain better image descriptors. We extend our proposal to multi-regional image representation and we combine our proposal with other spatial weighting measures. The descriptors derived from the salient regions improve the performance in three well known image retrieval datasets as we show in the experiments.