Image Processing Pipeline MCQ
Color spaces, geometric transformations, convolution filtering, and edge detection for classical computer vision preprocessing.
Color Spaces MCQ
Color spaces in Computer Vision
Choosing a color space affects segmentation thresholds, augmentation, and learning. RGB is natural for capture and display; HSV decouples hue; LAB improves perceptual distances; YCbCr separates luma for compression.
Why convert?
Different tasks need invariances: lighting robustness, perceptual uniformity, or compatibility with codecs and print.
Quick reference
Additive vs subtractive
RGB adds light; CMYK subtracts with inks—different gamuts and workflows.
Chroma subsampling
JPEG/video often store color at lower resolution than luma—saves bits with small perceived loss.
Gamma & linear
Compositing and physically based steps may need linear light; display encoding is nonlinear.
LAB distance
ΔE-style distances in LAB better match human judgment than raw RGB Euclidean distance.
Typical conversions
Camera RGB → (white balance) → sRGB → optional HSV/LAB/YCbCr for algorithm
Image Transformations MCQ
Geometric transformations
Aligning templates, augmenting datasets, rectifying documents, and stitching panoramas all rely on knowing how coordinates map under linear and projective models.
Affine vs homography
Affine preserves parallelism; homography models plane-to-plane perspective and can rectify quadrilaterals to rectangles.
Essentials
Interpolation
Nearest, bilinear, bicubic trade quality vs speed when resampling warped coordinates.
Composition
Order of rotation and translation matters; use homogeneous matrices to chain transforms.
Downsampling
Prefilter before shrink to limit aliasing—same Nyquist intuition as sampling theory.
Inverse mapping
For each destination pixel, sample source at inverse-warped location to avoid holes.
Transform hierarchy
Translation ⊂ Rigid ⊂ Similarity ⊂ Affine ⊂ Projective
Image Filtering MCQ
Image filtering fundamentals
Filtering builds almost everything downstream: denoise before edge detection, build pyramids, or preprocess for feature extractors. Know linear vs nonlinear behavior and border policies.
Convolution intuition
Slide a template over the image, sum weighted neighbors—implements low-pass, high-pass, or matched filters depending on kernel values.
Building blocks
Low-pass
Gaussian and box filters suppress noise and fine texture—also used as baseline for unsharp masking.
High-pass / edges
Derivative kernels emphasize changes; larger σ blur-first-then-derive stabilizes noisy derivatives.
Nonlinear
Median, bilateral, and morphological filters handle outliers and preserve structure differently than convolutions.
Separable Gaussian
Two 1D passes implement 2D Gaussian efficiently—critical for real-time pipelines.
Typical preprocessing
Capture → Denoise / normalize → Filter bank or CNN layers → Tasks
Edge Detection MCQ
Edge detection in Computer Vision
Edges mark intensity discontinuities—often object boundaries. Classical pipelines combine smoothing, gradient estimation, thinning, and linking.
Canny highlights
Gaussian pre-smoothing, gradient magnitude/direction, non-maximum suppression along normal, hysteresis to trace strong edges with weak continuity.
Ideas to remember
First derivatives
Sobel/Prewitt approximate ∂I/∂x and ∂I/∂y; magnitude combines both; direction matters for NMS.
Noise
Derivatives amplify noise—blur σ trades edge localization vs robustness.
Second derivatives
Laplacian zero-crossings locate edges but are sensitive to noise without careful scaling.
Linking
Hysteresis uses high/low thresholds to reduce streaking while preserving weak edge segments attached to strong ones.
Typical Canny flow
Smooth → Gradients → Magnitude/angle → NMS → Hysteresis