Segmentation Basics — Interview Q&A

Answer: Classifying pixels as foreground vs background by comparing intensity to one or more thresholds—produces binary or multi-label masks.

Answer: Single threshold T for the whole image: pixel → foreground if I > T (or < depending on type). Fast but fails under uneven lighting or overlapping histograms.

Answer: When objects are darker than background (or you want white objects on black mask). Complement of standard binary THRESH_BINARY.

Answer: Assumes roughly bimodal histogram; picks T that minimizes intra-class variance (equivalently maximizes between-class variance). Automatic, no manual T.

_, bw = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

Answer: Unimodal or flat histograms, uneven illumination, low contrast, or when foreground fraction is tiny—histogram may not have clear valleys.

Answer: For each pixel, threshold = mean of local neighborhood − C. Handles varying illumination; needs block size larger than foreground features.

Answer: Local threshold from Gaussian-weighted mean instead of flat mean—smoother local estimates, slightly better on gradual shading.

Answer: Odd window size defining local neighborhood. Too small: noisy mask; too large: loses detail near object boundaries.

Answer: Local methods using mean and standard deviation to set threshold—good for document and degraded scans with uneven background.

Answer: Homomorphic filtering, background normalization, CLAHE on luminance, or large-kernel low-pass estimate of illumination to flatten.

Answer: Often convert to HSV and threshold H/S/V ranges (e.g. colored ball)—more robust than RGB for hue-based objects under some lighting.

Answer: Several thresholds to get multiple classes (e.g. tissue types). Extension of Otsu exists (multi-Otsu) but cost grows with levels.

Answer: Manual inspection, ROC on validation set, entropy methods, or trial with domain constraints (known object brightness).

Answer: Adaptive threshold or Sauvola-class; deskew/denoise first; morphology to clean speckles—DL methods also used for hard cases.

Answer: Sensor noise, shadows, partial overlap of histograms—use morphology, median blur pre-threshold, or adaptive methods.

Answer: Yes—opening removes pepper noise, closing fills small holes in foreground; preserves label if structuring element smaller than features.

Answer: For complex scenes, semantic segmentation wins; classical thresholding remains fast for controlled lighting, industrial vision, and documents.

Answer: Shrinks coefficients toward zero—used in denoising, not classical image binarization. Mention only if interviewer asks denoising context.

Answer: Same logic but ensure consistent range ([0,1] vs 0–255). Always know your image dtype before comparing to T.

Answer: Often denoise/blur lightly → threshold → morphological cleanup—order depends on whether blur destroys thin structures.

Answer: Non-linear image ops based on set theory with a structuring element—shape analysis, denoising binary masks, and extracting boundaries.

Answer: SE fits inside foreground: output pixel on only if SE fully contained in foreground—shrinks objects, removes thin protrusions, separates touching objects if SE sized right.

Answer: SE hits foreground: output on if SE overlaps foreground—grows objects, fills small holes and gaps, reconnects broken strokes.

Answer: Erosion then dilation with same SE—removes small bright noise (pepper) and smooths boundaries without changing coarse size as much as raw erosion.

Answer: Dilation then erosion—fills small dark holes (salt holes in foreground), connects nearby components, smooths inward concavities.

Answer: Small binary mask defining neighborhood shape—disk is isotropic; rectangle aligns to axes; size controls scale of effect.

Answer: Disk gives isotropic rounding; axis-aligned square can preserve Manhattan geometry—choice affects anisotropy of shrink/grow.

Answer: Erosion of foreground equals dilation of background complement (with reflected SE)—lets derive properties and implement efficiently.

Answer: Boundary ≈ original − eroded (or XOR variants)—gives outer contour ring depending on SE.

Answer: Dilation − erosion—edge strength similar to gradient magnitude on binary shapes; outer/inner gradients are variants.

Answer: Image − opening—extracts bright small details larger than SE removed by opening; black top-hat is closing − image for dark details.

Answer: Template matching for specific pixel configurations—detects corners, endpoints, pruned skeletons using paired SE foreground/background patterns.

Answer: Min/max filter over SE neighborhood—useful for texture and background estimation; different from binary set interpretation but analogous.

Answer: No—generally opening∘closing ≠ closing∘opening. Order matters for pipeline design.

Answer: Opening removes speckle; closing fills letter breaks—tune SE smaller than stroke width to avoid destroying characters.

Answer: Removes/fills larger features; too large destroys valid structure—scale SE to minimum noise size you want removed.

Answer: Geodesic reconstruction, conditional dilation from border, or flood-fill tricks—classic interview “beyond closing” answer.

Answer: Reduce foreground to 1-pixel-wide medial axis preserving topology—sequential morphological thinning or distance-transform methods.

Answer: MORPH_OPEN, CLOSE, GRADIENT, TOPHAT, BLACKHAT, HITMISS—single API with MORPH_ERODE/DILATE base.

kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
clean = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel)

Answer: Morphology is min/max (nonlinear); convolution is weighted sum. Morphology preserves sharp extrema semantics for shapes.