Unsupervised Cross-Domain Deep-Fused Feature Descriptor for Efficient Image Retrieval

Deep feature-based methods show advantages in image retrieval, yet their robustness and generalization warrant further investigation. Toward this end, we propose a robust unsupervised cross-domain deep-fused feature descriptor (CDFD) method for efficient image retrieval. It analyzes deep features from a frequency domain perspective, employing discrete wavelet transform (DWT) and discrete cosine transform (DCT) to enhance its robustness and distinguishability. Specifically, we begin with a distinct perspective and introduce DWT to identify low-frequency and high-frequency components in deep features. A multitrait spatial fusion strategy is proposed to integrate various features and strengthen low-frequency information. It can highlight the target region in the image and suppress the background clutters by introducing a low-pass filter, thus improving the discriminating ability of the feature representation. To enhance the robustness of the features, a cross-domain convergence scheme is proposed, which enables the rational integration of spatial and frequency-domain information by utilizing DCT. In addition, a computationally simple decrement query expansion technique is introduced. Using it with CDFD can effectively improve retrieval performance. Extensive comparative experiments on several benchmark datasets demonstrate that our CDFD method is robust and effective in image retrieval and outperforms several unsupervised state-of-the-art methods. The source code is published at https://github.com/sevenjava/CDFD.