WSG-FEC: Gray-Box Frequency-Domain Explanation of CNN Models Based on Weakly Supervised Learning

Convolutional neural networks (CNNs) have achieved remarkable progress in recent years, largely driven by advances in computational hardware. However, their increasingly complex architectures continue to pose significant challenges for interpretability. Existing explanation methods predominantly rely on spatial saliency representations and therefore fail to capture the intrinsic frequency-domain characteristics of image data. This work introduces WSG-FEC, a Weakly Supervised Gray-box Frequency-domain Explanation framework that integrates multilevel Haar wavelet decomposition with gradient-weighted class activation mapping (Grad-CAM) to generate frequency-aware explanations. By leveraging hierarchical wavelet structures, the proposed method produces frequential–spatial saliency maps that overcome the low-resolution limitations of white-box approaches and the high computational cost of black-box perturbation methods. Quantitative evaluations using insertion and deletion metrics demonstrate that WSG-FEC provides more detailed, efficient, and interpretable explanations, offering a novel perspective for understanding CNN decision mechanisms.