Image Processing Pipeline — Interview Q&A

Answer: A coordinate system for representing colors as numeric tuples (e.g. three numbers for trichromatic display). Different spaces emphasize different properties—device RGB for screens, HSV for intuitive hue/saturation edits, LAB for perceptual distance.

Answer: Red, Green, Blue primary lights added for displays. Each channel 0–255 (8-bit) combines to reproduce colors on monitors. It is device-dependent unless tied to a standard like sRGB.

import cv2
hsv = cv2.cvtColor(bgr, cv2.COLOR_BGR2HSV)

Answer: Libraries like OpenCV store channels as B, G, R. Algorithms are identical if consistent, but visualization and pre-trained weights expecting RGB need an explicit swap.

Answer: Hue (color tint on a wheel), Saturation (colorfulness vs gray), Value/Brightness (intensity). Cylindrical geometry separates chromatic from achromatic changes more intuitively than RGB for some tasks.

Answer: Segmenting by hue ranges (e.g. colored objects), thresholding saturation/value to ignore shadows differently than RGB splits, and some augmentations that tweak hue/saturation while preserving identity.

Answer: L* is lightness; a*, b* are color-opponent dimensions. Euclidean distance in LAB approximates perceptual difference better than RGB. Useful for color transfer, quality metrics, and some clustering tasks.

Answer: Separates luma (Y) from chrominance (Cb, Cr). Used in JPEG and video codecs because human vision is more sensitive to brightness than color—enabling chroma subsampling.

Answer: Subtractive printing (cyan, magenta, yellow, key/black). Less common in core CV training; relevant for print QA, packaging inspection, and prepress—not for typical RGB camera pipelines.

Answer: It is a single-channel intensity representation, often derived from RGB via weighted sum. It discards chrominance—good for edge detection and speed when color is irrelevant.

Answer: Sensors measure roughly linear light; displays apply gamma encoding (sRGB transfer function) for perceptual uniformity. Some photometric algorithms (deblur, relighting) need linearization via inverse gamma before physical correctness.

Answer: Nonlinear mapping between stored values and displayed intensity to match human brightness perception and legacy CRT behavior. Applying gamma wrong can break color statistics and blur/threshold results.

Answer: The range of colors a device or space can represent. Wide-gamut displays (P3) vs sRGB differ; out-of-gamut colors clip or map when converting—important for medical imaging and professional color.

Answer: Reference neutral light (e.g. D65) for interpreting RGB values. Different cameras/AWB change apparent colors; robust pipelines account for illumination via white balance or learning.

Answer: Full luma resolution but quarter resolution for chroma planes—exploits lower acuity for color. Can cause color fringing on sharp edges when decoded; relevant for video compression pipelines.

Answer: Sometimes for model input (zero mean / unit var per channel). For photometric consistency, consider normalization that preserves color ratios—or work in a space suited to the task (e.g. LAB L channel only).

Answer: Random brightness/contrast often in RGB or HSV; hue jitter in HSV. Extreme hue shifts may leave gamut or break class semantics—keep augmentations label-safe.

Answer: RGB thresholding needs rules in 3D (ranges per channel or distance to a color). HSV can separate hue cone from lighting via S/V gating—still not perfect under colored illumination.

Answer: Applying independently to R,G,B shifts colors (color cast). Often convert to LAB and equalize L only, or use CLAHE on luminance to preserve chroma.

Answer: Retinex-style ideas, white balance, homomorphic filtering (separate illumination/reflectance in log domain), or learning-based methods. Interviews reward naming tradeoffs (artifacts vs compute).

Answer: Often: load image → ensure correct color order → optional WB/gamma fix → resize/crop with good interpolation → normalize to tensor. Order matters: resize after linearization for photometric tasks; many DL pipelines keep it simple in sRGB uint8.

Answer: A mapping that moves pixel locations—translation, rotation, scale, affine, or perspective—while optionally resampling intensities. It changes spatial layout but not the semantic label if the transform is label-consistent (e.g. bbox corners transformed too).

Answer: Shifting all pixels by offsets (tx, ty). Implemented by moving the sampling grid or adjusting the transform matrix with identity + translation column. Boundaries may require padding or cropping.

Answer: Isotropic: same scale sx = sy preserves angles. Anisotropic: sx ≠ sy stretches content—can turn circles into ellipses. Know effect on aspect ratio for detection labels.

Answer: Linear part is matrix [[cos θ, -sin θ],[sin θ, cos θ]]. In practice pick a rotation center (image center) via translate-rotate-translate composition. Large rotations need bigger canvas or cropping.

Answer: Horizontal flip is a common label-preserving augmentation for many object classes; vertical flip may break semantics (people, text, traffic scenes). Always validate against dataset semantics.

Answer: Represent point (x,y) as (x,y,1). Allows affine maps as 3×3 matrices acting on homogeneous vectors, unifying translation with linear maps for composition.

Answer: Maps parallel lines to parallel lines: combination of linear transform and translation—rotation, scale, shear. Preserves ratios along lines but not necessarily lengths or angles unless constrained (similarity/euclidean).

Answer: Six (4 in the 2×2 linear part + 2 translation). You need 3 point correspondences (non-degenerate) to estimate it in general.

Answer: Projective maps preserve collinearity but not parallelism—parallel world lines can converge in the image (vanishing points). Needed for planes viewed at an angle, document scanning, and bird’s-eye view from ground cameras.

Answer: A 3×3 projective transform (up to scale) mapping one plane to another in pinhole imaging. Relates two views of the same planar surface. Estimated from 4 point correspondences (DLT) with constraints.

Answer: Forward: map source→dest can leave holes and overlaps. Inverse: for each destination pixel, sample source via inverse map—avoids gaps and is standard in OpenCV warp* with a chosen interpolator.

Answer: Mapped coordinates land between pixels. Nearest, bilinear, bicubic choose neighborhood weights—trade speed vs aliasing/blur. Downscaling may need prefiltering to avoid aliasing.

import cv2
M = cv2.getRotationMatrix2D((cx, cy), angle, scale)
out = cv2.warpAffine(img, M, (w, h))

Answer: Rotation/scale can push content outside the original canvas—either expand canvas with padding (constant, reflect) or crop to a fixed size. Detection boxes must be clipped or transformed consistently.

Answer: Apply the same spatial map to image and mask (nearest-neighbor interpolation for label masks to avoid fractional classes). For instance segmentation, warp polygons or rasterize after transform.

Answer: Aligning two images of the same scene into a common coordinate frame—via feature matching + homography/affine, optical flow, or optimization. Used in medical imaging, panorama stitching, and super-resolution.

Answer: Rotation + uniform scale + translation (4 DOF in 2D). Preserves angles and ratios of lengths—good model when perspective effects are weak.

Answer: Rotation + translation only—preserves distances and angles (3 DOF in 2D). Models camera motion parallel to the plane or object pose without scale change.

Answer: Multiply their homogeneous matrices in application order (rightmost often applied first to a column vector—be consistent with your library convention).

Answer: warpAffine uses a 2×3 affine map; warpPerspective uses full 3×3 homography. Choose based on whether parallelism must be preserved (affine) or full perspective correction is needed.

Answer: No—radial/tangential distortion is nonlinear and modeled separately (Brown-Conrady) before or jointly with pinhole projection. Undistort first, then apply homography for many planar AR/document pipelines.

Answer: Computing a new image where each output pixel is a function of a neighborhood of input pixels. Linear filters use weighted sums (convolution/correlation); nonlinear filters include median, bilateral, morphological ops.

Answer: Slide a kernel over the image; at each location, sum of elementwise products of kernel and flipped neighborhood (strict convolution). Many libraries implement cross-correlation without flip—know which convention your framework uses.

Answer: For symmetric kernels (Gaussian), results match. For asymmetric kernels (Sobel direction), flip matters for strict signal-processing convolution vs correlation.

Answer: Small matrix of weights defining neighborhood contributions. Size (e.g. 3×3, 5×5) sets spatial support; larger kernels increase blur radius and compute cost (~kernel area per pixel).

Answer: Median (order-statistic, good for salt-and-pepper), bilateral (edge-preserving smoothing), morphology (min/max). They do not obey superposition like convolutions.

Answer: Smooth low-pass filtering that reduces noise and high frequencies while avoiding sharp ringing like ideal low-pass. Separable implementation makes it fast; σ controls blur strength.

import cv2
blur = cv2.GaussianBlur(img, (5, 5), sigmaX=1.0)

Answer: A 2D kernel K can equal outer product v vᵀ. Convolving with K is equivalent to 1D conv along rows then columns—cost drops from O(WHk²) to O(2WHk) for k×k support.

Answer: Simple average of neighborhood—fast (especially with integral images) but has a sharp frequency nulls profile vs Gaussian; can create blocky artifacts compared to Gaussian blur.

Answer: Impulsive salt-and-pepper noise where mean blur smears outliers. Median preserves edges better than Gaussian for that noise but is costlier and can remove thin structures.

Answer: Weighted average where weights drop with both spatial distance and intensity difference—smooths flat regions but preserves sharp edges. Used for denoising and tone mapping; slower than Gaussian.

Answer: Extends the image border so output size can match input (same padding) or follow strict convolution (valid). Modes: zero, reflect, replicate, wrap—choice affects edges and CNN behavior.

Answer: Neighborhoods at borders are incomplete; padding synthesizes missing neighbors. Wrong padding can cause dark/bright fringes—noticeable on small images and CNN feature maps.

Answer: Emphasize high frequencies by adding a scaled Laplacian-like response or subtracting a blurred version from the original—makes edges pop but can amplify noise.

Answer: Enhancement: original + amount × (original − blurred). The difference is a high-boost of details; used in photography and preprocessing (with care for noise).

Answer: Stride >1 subsamples the output—like convolve-then-downsample. Larger stride increases receptive field progression and reduces spatial size; different from stride-1 spatial filtering used in preprocessing.

Answer: Gaussian in space ↔ Gaussian in frequency; it attenuates high frequencies smoothly. Helps before subsampling to limit aliasing (Nyquist)—ties back to image basics.

Answer: Gaussian noise: linear smoothing (Gaussian blur) or Wiener/BM3D-class methods at higher level. Salt-and-pepper: median or morphological openings/closings.

Answer: Finite differences (Sobel/Prewitt) are short convolution kernels approximating gradients—high-pass. Often paired with prior Gaussian smoothing to reduce noise before edge detection.

Answer: So the DC gain is 1—preserves average brightness. Unnormalized Gaussian sums to 1 after discretization normalization; forgetting normalization scales image intensity.

Answer: GaussianBlur builds separable Gaussian internally. filter2D applies arbitrary kernel (correlation-style in OpenCV)—flexible for custom linear filters.

Answer: Finding boundaries where intensity changes rapidly—object outlines, surface markings, shadows. Edges are local; full segmentation groups pixels into regions.

Answer: Vector of partial derivatives (Ix, Iy). Magnitude ‖∇I‖ shows edge strength; direction is perpendicular to the edge (along max rate of change).

Answer: Discrete 3×3 separable approximation of derivatives with slight smoothing (center weight). Gx and Gy kernels estimate Ix, Iy; combine for magnitude and angle.

Answer: Similar 3×3 derivative masks; weights differ slightly (Sobel emphasizes center more). Both approximate first derivatives; results are often close for interviews.

Answer: Second derivative—zero-crossings align with edges. Sensitive to noise; often applied to Gaussian-smoothed image (LoG) for stability.

Answer: Locations where Laplacian of Gaussian changes sign—candidate edges. Need additional filtering to reduce spurious responses from noise.

Answer: Differentiation amplifies noise. Gaussian low-pass reduces noise while keeping meaningful discontinuities; leads to LoG or smooth gradients for Canny.

Answer: 1) Gaussian smooth 2) gradient magnitude/direction 3) non-max suppression 4) hysteresis with high/low thresholds to link strong edges and reject weak noise.

Answer: Thins edges: at each pixel keep magnitude only if it is a local max along the gradient direction—produces one-pixel-wide ridges.

Answer: Use high T_hi to accept strong edges, low T_lo to continue chains from strong pixels—reduces broken edges while suppressing isolated weak noise.

Answer: Thick edges blur object boundaries, hurt subpixel localization, and complicate linking. NMS aims for single-pixel thickness.

Answer: Options: max/mean gradient across channels, convert to luminance first, or vector gradient methods. Channel-wise max is simple; luminance can discard chromatic edges.

Answer: Creates spurious large gradients. Median filter first can help; Gaussian blur before derivatives is standard for Gaussian noise.

Answer: Large σ: fewer, smoother edges (coarse structure). Small σ: more detail and noise. Multi-scale edge detection combines responses at several σ.

Answer: Smooth with Gaussian, apply Laplacian, find zero-crossings—Marr-Hildreth approach. Approximated by Difference of Gaussians (DoG) in some pipelines.

Answer: Connecting broken edge pixels into curves using proximity, direction continuity, or global optimization—step after local edge detection for contours.

Answer: threshold1, threshold2 for hysteresis (low/high), apertureSize for Sobel, L2gradient flag for magnitude formula. Tune for your noise and scale.

import cv2
e = cv2.Canny(img, 50, 150)

Answer: Gradient points in direction of steepest ascent; edge normal is often aligned with gradient; edge tangent is perpendicular.

Answer: Fit parabola to gradient magnitudes along normal, moment-based refinement, or optimization—used in metrology and calibration.

Answer: Edges are local discontinuities; segmentation assigns each pixel to a region/object. Edges can guide segmentation (watershed, active contours, graphs).