Convolutional Neural Network-based image retrieval with degraded sample

Over a decade, convolutional neural networks (CNNs) have been applied extensively on various tasks related to images. Given an input image, a CNN model will investigate the content and deduce the representation of this image using a model's structure built from hidden neurons. This representation analyzes data semantically, which helps to solve semantic issues, such as image retrieval. To verify the above viewpoint, this study addresses the problem of using features learned from a CNN model to perform image retrieval. To more emphasize the efficiency of learned features, we consider degraded images and their enhanced version as queries and search for similar ones in the gallery set. Data augmentation is also applied to increase the number of images in the gallery. The experiments are conducted on a multi-view dataset, smallNorb. Experimental results are reported both in quantity and quality.