Image Quality: Noise, Contrast, and Structure#

The quality of a plate image directly determines how well detectors and enhancers can perform. PhenoTypic’s plot.diagnostics() method provides quantitative metrics for the three dimensions of image quality.

Noise#

Noise is random pixel-level variation that is not part of the true image signal. It introduces false texture in the agar background and blurs colony boundaries.

Sources: Sensor read noise, shot noise (photon counting), thermal noise (long exposures), compression artifacts.

Metrics:

  • SNR (Signal-to-Noise Ratio) — ratio of mean signal to noise standard deviation. Higher is better. Values below 10 indicate significant noise.

  • Correlation length — the spatial extent of correlated noise. Short correlation (a few pixels) indicates random noise; long correlation suggests structured artifacts (e.g., banding, uneven illumination).

Mitigation: GaussianBlur, MedianFilter, StableDenoise (BM3D), VisuShrinkEnhancer (wavelet).

Contrast#

Contrast measures how well colonies separate from background in intensity.

Sources of low contrast: Under-exposure, faint pigmentation (early time points), agar that matches colony color, condensation on plate lid.

Metrics:

  • RMS contrast — root-mean-square intensity deviation. Low values mean the histogram is narrow (everything looks similar).

  • Michelson contrast — (Imax − Imin) / (Imax + Imin). Values near 1.0 indicate strong separation; values near 0 mean colonies and background have similar brightness.

  • Dynamic range — fraction of the theoretical bit depth in use. A 16-bit image that only uses values 0–1000 has very low dynamic range.

Mitigation: CLAHE, ContrastStretching, HomomorphicFilter.

Structure#

Structure metrics assess the spatial organization of intensity patterns — how sharp the colony edges are and how consistent the spatial layout is.

Metrics:

  • Gradient mean — average edge strength across the image. Higher values mean sharper colony boundaries, which makes threshold-based detection easier.

  • Coherence — consistency of gradient orientation. High coherence on a grid plate suggests well-organized, regularly spaced colonies.

Mitigation: UnsharpMask, SobelFilter (enhance edges before detection).

Quality-Guided Pipeline Design#

Assess image quality before choosing enhancers:

  1. High noise, adequate contrast → denoise first, then detect

  2. Low noise, low contrast → contrast enhancement, then detect

  3. Uneven illumination → HomomorphicFilter before anything else

  4. Soft edges → UnsharpMask or edge-enhancing filter

  5. All metrics good → skip enhancement, detect directly