Abstract
Deep learning techniques can be used to describe image content, which is a good way to reduce the semantic gap between low-level and high-level features. However, convolutional neural networks (CNNs) exhibit texture bias and largely ignore global object shape. This does not conform to human perception and can fail to gain the advantages of both low-level features and deep features. To address this problem, in this work, the main focus was shifted from using simple deep features to the use of sublimated deep features, which incorporate global object shape and color features. Along these lines, in this work, a novel image retrieval method named the sublimated deep feature histogram (SDFH) was proposed. The main highlights are: 1) An effect orientation feature was introduced, namely, orientation-selective feature, to mimic the orientation-selection mechanism. This provides a good representation of global object shape and reduces the side-effects of texture bias. 2) A new concept was developed, namely color perceptual feature, to address the shortcoming of deep features-that they discard color features. This includes color cues in the deep features and provides a more discriminating representation. 3) The orientation selective and color perception mechanisms were effectively mimicked to provide a compact yet efficient representation, and propose an effective transfer learning method called gain whitening learning. By carrying out comparative experiments, it was demonstrated that sublimated deep features can provide highly competitive retrieval performance (in terms of mean average precision) using a pre-trained CNN model applied to well-known benchmark datasets. Our results provide new insights into image retrieval based on the mechanisms of the primary visual cortex (V1). Furthermore, the method is more in line with human perception than other methods.