PCANet-based processing of MAE latent representations for CME region localization. The latent representations (F) are processed by two-stage unsupervised PCA filters to extract multidirectional features, which are binarized and aggregated into histograms (f_i) for robustness. A subsequent energy function integrates these responses into a saliency map, upsampled to 512 × 512, to generate the final CME colocalization map (brighter for higher CME probability).