Image retrieval using underlying importance feature histogram

Deep features can exhibit superior retrieval performance than low-level features. However, low-level features (e.g. colour and orientation) can be extracted by generally imitating the human visual perceptual system. Combining human-like low-level and deep features can harmoniously yield more discriminative representations. However, it remains challenging. To address this problem, a new representation method for image retrieval, namely the underlying importance feature histogram (UIFH), is presented in this study. Its main highlights are: (1) This new method extracts low-level features by simulating the human visual perception mechanism, such as opponent colour and orientation selectivity mechanisms. (2) Inspired by the salience evaluation mechanism, the new method can harmoniously evaluate the underlying importance information between deep and low-level features. (3) Assisting the various important information can facilitate the UIFH. It can substantially improve the discriminative power of representation. Comprehensive experiments on seven benchmark datasets demonstrated that the proposed UIFH method outperforms some recent state-of-the-art methods based on pre-trained models. The proposed UIFH method is suitable for the retrieval scenes where images have various colours and prominent orientations.