Research
Deep Phi is built on published research: structural masking, semantic-free ROI extraction, structure-constrained super-resolution and enhancement. The papers share protocols and results — never φ's formula.
The masks stay stable across a wide range of natural images and acquisition conditions, and capture meaningful structure beyond classical edge- or gradient-based representations — a generic guidance signal for downstream tasks.
On 120 natural images (Flickr2K) and the UETT4k-Anti-UAV set (30 images), it delivers 95–99% spatial reduction, outperforms Sobel/Laplacian baselines and matches Shannon-entropy approaches — with higher spatial coherence.
The multi-scale version yields more compact masks, fewer spurious regions and significantly more stable ROI selection across window sizes than single-scale formulations — a robust canonical extension.
On a single-step ×4 model over 223 Flickr2K images, Φ-conditioning consistently improves reconstruction over classical interpolation baselines, acting as an interpretable structural prior — most of all in regions of high structural complexity.
Across 200 natural images, 15 methods and 4 window sizes, the multi-scale φ-guided CLAHE variant surpasses, at its optimal scale, standard CLAHE by 73.4%, guided CLAHE by 47.9% and Sobel-guided CLAHE by 6.2%.
The φ signal keeps a positive correlation to its reference across every degradation regime; on real Canon/Nikon captures PHI-SR shows a perceptual advantage (LPIPS) over LIIF — including under added degradation, and markedly so on thermal infrared.
Across two visuomotor environments: up to −75% prediction loss on clean data; planning success under noise lifted from 28% to 44% (7 of 25 versus 11 of 25 — not significant at this trial count, p = 0.377); on noisy manipulation the baseline rises from 0% to 4–8%. The clean-manipulation effect stays inconclusive.
The publications describe behaviour, protocols and results. φ's internal formula is not disclosed — by design.