Approximate Multidimensional Discrete Cosine Transform-Based Deep Attention Network for Cross-Spectral Periocular Recognition

This article introduces a novel, parameter-free 3-D attention mechanism that utilizes the concept of approximate multidimensional discrete cosine transform (AM-DCT) for recognizing individuals using cross-spectral periocular image modalities. The proposed attention mechanism simultaneously considers the entire convolutional feature space to measure the significance of each feature location. The AM-DCT attention module effectively establishes a correlation between the importance of 3-D convolutional features and their impact on the energy distribution of the transformed feature map. The utilization of multiplierless discrete cosine transform (DCT) approximations across multiple dimensions results in a reduction of computational overhead for the new attention module. The feature reconstruction process involves preserving only significant transform coefficients and discarding the insignificant ones according to an energy thresholding criterion. The efficacy of the proposed deep attention network is assessed using five cross-spectral periocular datasets comprising a large number of unconstrained images captured in visible, near-infrared, and night vision spectra. The experimental results demonstrate that the proposed network efficiently achieves state-of-the-art recognition performances on all datasets with no additional parameter overhead and minimal training and inference time. Extensive ablation studies are conducted to validate the significance of the proposed AM-DCT attention mechanism for cross-spectral periocular recognition.