Expand description
Tier: CnnLight — lightweight learned SR (FSRCNN / ESPCN).
Permissive weight source (gated ①, see docs/plans/gated-av-jobs-2026.md §2,
row “SR CNN-Light”): FSRCNN / ESPCN from the OpenCV zoo (Apache-2.0),
dropped into vendor/vision/sr/{fsrcnn,espcn}/ and converted to ONNX.
Pipeline (the classic FSRCNN/ESPCN convention): these networks super-resolve
only the luma (Y) channel. We convert RGB8 → YCbCr (BT.601), run the ONNX
model on a single-channel 1×1×H×W Y tensor normalised to [0,1], upscale
the chroma (Cb/Cr) classically with bilinear interpolation, then recombine
YCbCr → RGB8 into out. This is non-generative (generative = false): it
recovers luminance detail without hallucinating texture, which is why it is
the honest Sharpen-class tier for medical/forensic use.
#[cfg(not(feature = "vision-onnx"))]→ fail closed (FeatureDisabled); never invents pixels.#[cfg(feature = "vision-onnx")]→ realortinference.
No unwrap() outside #[cfg(test)].
Constants§
- BACKEND_
ID - Backend id reported for this tier.
Functions§
- cnn_
light_ super_ resolve - Super-resolve
rgbbyscaleusing a CnnLight (FSRCNN/ESPCN) ONNX weight.