Adaptive homomorphic filtering with a compact neural parameter predictor for illumination-robust image enhancement

Homomorphic filtering separates illumination from reflectance in the logarithmic frequency domain and remains one of the most economical ways to correct uneven lighting. Its practical weakness is that the shape of the emphasis filter must be tuned by hand for every image. This paper makes three contributions. First, it shows that the classical four-parameter Gaussian emphasis filter is structurally over-parameterised: the sharpness constant and the cut-off radius enter the transfer function only through a single ratio, so they cannot be estimated separately from any observation. Collapsing them into one effective cut-off radius yields an identifiable three-parameter filter. Second, it proposes AHF-Net, in which a compact multilayer perceptron of about one thousand weights maps a fourteen-dimensional illumination descriptor of the observed image onto the filter parameters and a local contrast limit, so that no manual tuning is required at run time. Third, it replaces per-channel processing by luminance-domain filtering with chroma-ratio recombination, which removes the colour drift produced by filtering the three primaries independently. On a benchmark of three hundred and forty images with known ground truth and evaluated under scene-grouped cross-validation, the method reaches 18.98 dB peak signal to noise ratio, 0.735 structural similarity and 12.68 mean colour difference, against 16.78 dB, 0.648 and 17.38 for the classical fixed-parameter filter. Colour difference improves by 3.38 units relative to per-channel filtering. Cases in which the method does not improve on existing approaches are reported and analysed.