Saral Shiksha Yojna
Courses/Computer Vision

Computer Vision

CSE471
Prof. Makarand Tapaswi + Prof. Charu Sharma•Spring 2025-26•4 credits
Sample Papers/Mock Paper 4 — Image Processing Foundations + ML Basics + Metrics

Mock Paper 4 — Image Processing Foundations + ML Basics + Metrics

Duration: 180 min • Max marks: 100

Section A — Short Answer (1-2 marks each, 20 marks)

20 marks
  1. 1.Response of the 3×3 Laplacian kernel [[0,-1,0],[-1,4,-1],[0,-1,0]] on a flat region (all 100s)?1 m
  2. 2.Apply 5×5 Gaussian σ=1 then 5×5 Gaussian σ=2. What single Gaussian produces an equivalent effect?2 m
  3. 3.The Fourier transform of a real image is conjugate symmetric: F(−u, −v) = F*(u, v). Storage consequence?1 m
  4. 4.A binary image is eroded 5× with a 3×3 box SE then dilated 5× with the same SE. Why is the result NOT the same as the original?2 m
  5. 5.Write Otsu's between-class variance and explain in one sentence why maximising it gives the optimal threshold.2 m
  6. 6.Gamma correction formula. What γ brightens dark regions?1 m
  7. 7.8-bit image with most pixels in 0-50. After histogram equalisation: (a) global dynamic range, (b) contrast in dark regions.2 m
  8. 8.Classification vs regression as ML tasks.1 m
  9. 9.Why is cross-entropy preferred over MSE for classification with softmax output?2 m
  10. 10.L1 vs L2 regularisation difference.1 m
  11. 11.Compute F1 given Precision = 0.8, Recall = 0.6.2 m
  12. 12.Confusion matrix: relate TP, FP, TN, FN.1 m
  13. 13.Given covariance matrix C with sorted eigenvectors v_1, …, v_d, how do you project a data point x onto the top-k principal components?2 m
  14. 14.k-means in two sentences.1 m

Section B — Conceptual / Explanation (4-6 marks each, 40 marks)

40 marks
  1. 1.State the Convolution Theorem, explain its practical significance for image processing, and describe a case where you'd choose frequency-domain filtering over spatial-domain filtering.5 m
  2. 2.Describe the bilateral filter. How does it differ from a Gaussian filter and why does it preserve edges?5 m
  3. 3.Explain the Hough transform for line detection in 5 steps. Why polar (ρ, θ) representation over slope-intercept y = mx + b?5 m
  4. 4.Compare k-NN classification with logistic regression for image classification. Two scenarios where each would be preferred.4 m
  5. 5.Derive the gradient of cross-entropy with softmax with respect to the logits. Show it simplifies to (softmax_output − target).6 m
  6. 6.Confusion matrix on 3-class problem: rows actual, cols predicted. [cat 50/8/2; dog 12/45/3; bird 5/7/68]. (a) Per-class precision, recall. (b) Macro-averaged F1.5 m
  7. 7.Self-driving model has 97% accuracy but misses pedestrians 40% of the time. Explain via class imbalance and propose two engineering solutions.5 m
  8. 8.Compare PCA and t-SNE for image-feature dimensionality reduction. When to use each?5 m

Section C — Long Form (10 marks each, 40 marks)

40 marks
  1. 1.Binary fingerprint image (8 fingerprints, salt-and-pepper noise, some impressions broken). (a) Design a morphological pipeline producing one connected component per fingerprint. (b) Connected-component labeling finds 11 instead of 8 — what went wrong and how to fix? (c) Suggest structuring element shape and sizes.10 m
  2. 2.PCA worked example. Points x_1 = (2, 1), x_2 = (4, 3), x_3 = (6, 5), x_4 = (8, 7). (a) Mean. (b) Centred data. (c) 2×2 covariance matrix (use N − 1). (d) Eigenvalues. (e) Principal eigenvector and projection of each point.10 m
  3. 3.Canny-style edge detection pipeline. (a) Five stages with mathematical detail. (b) Why non-max suppression is needed and how it works. (c) Hysteresis vs simple thresholding.10 m
  4. 4.PCB visual quality inspection: < 100 ms/board, ≥ 95% defect detection, < 1% FP, adaptable to new designs. (a) Traditional CV vs deep learning — justify. (b) Architecture: backbone, heads, post-processing. (c) Adapt to a new PCB design with only 100 labelled examples.10 m

Track your attempt locally — score and time are recorded in your browser. (Coming soon: timed-attempt mode.)