Image retrieval via learning content-based deep quality model towards big data

Abstract

Image retrieval aims to search specific image from large-scale datasets. Traditional text-based and content-based image retrieval approaches have shown competitive performance. However, both of which are limited by semantic gap, i.e., they cannot reflect human perception of images. To narrow semantic gap in image retrieval, this paper proposes a deep neural network (DNN) based image retrieval method, where saliency map is derived to form human gaze shifting paths by constraint metrics. More specifically, we first design a DNN-based image saliency prediction. Subsequently, we leverage image quality assessment (IQA) algorithm to select high-quality salient regions, which will be concatenated in sequence by using proposed constraint metrics to mimic human visual perception. Afterwards, we leverage the CNN-based architecture for deep representation acquisition of each images, where spatial structure among salient regions can be well preserved. Subsequently, based on the quality score of the query image, a series of candidate images whose quality scores are similar to that of the query image are derived. Finally, we engineer a ranking distance metric to refine the candidate images to achieve image retrieval. Extend experiments demonstrate that our method outperforms several state-of-the-art algorithms.