CV Chapter 2 Image Processing Basics
Questions about the lecture 'Computer Vision' of the RWTH Aachen Chapter 2 Image Processing Basics
Questions about the lecture 'Computer Vision' of the RWTH Aachen Chapter 2 Image Processing Basics
-
- 1 / 106
-
Flashcards
What are the criteria for edge detection? [3]
1. Good detection with finding true and avoiding false edges (noise)
2. Good localization with detected close to true edges
3. Single response with returning one for each true edge point
What are the primary steps of edge detection? [3]
1. Smoothing
2. Filter for contrast which enhances edges
3. Localize edges with finding true maxima and thresholding thinning
What holds for smoothing for edge detection?
Gaussian influences structures with small/large value detects fine features/larger edges
What are the characteristics of the canny edge detector? [3]
1. Widely used in computer vision
2. Step-edges are corrupted by additive gaussian noise
3. First derivative approximates the optimizing operator signal-to-noise*localization
What are the characteristics of the non-maximum suppression of the canny edge detector? [2]
1. Find local maximum along gradient direction
2. Problem might be thinning with removing edges
What are the characteristics of the hysteresis thresholding of the canny edge detector? [2]
1. Use k_high (start edge chain) and k_low (continue edge chain)
2. Ratio k_high/k_low = 2
What is the MATLAB code for gaussian smoothing? [6]
>> for sigma=1:3:10
>> h = fspecial('gaussian‘, fsize, sigma)
>> out = imfilter(im, h);
>> imshow(out);
>> pause;
>> end
What is the MATLAB code for canny edge detector? [2]
>> edge(image,‘canny’);
>> help edge
What types of fitting as structure extraction exist? [2]
1. Template matching and 2. parametric search
What are the characteristics of template matching in fitting? [2]
1. Associate model (line, circle, shape) with observed features // Similar as template matching with edge templates (Traffic sign detection)
2. Correlation does not work because of value zero if one pixel off -> Use chamfer distance, distance transform
What are the characteristics of chamfer distance? [4]
1. Average distance to nearest edge pixel
2. D_Chamfer(x,y) = 1/T * Sum_[T(u,v)=1] dt(x+u,y+v)
3. More robust
4. Replace Kernel T by fixed number of points T
What are the advantages of chamfer distance? [3]
1. Fast and simple
2. Matches upright shapes with little intraclass variation
3. Matches candidate in longer recognition pipeline
What are the disadvantages of chamfer distance? [3]
1. Not discriminative and needs further verification
2. Many false positives detections for cluttered regions
3. Expensive for rotated rescaled objects
What is the definition of distance transform in 1D? [3]
1. Initialize with 0 and inf
2. Forward taking min(self, left+1) // Missing first
3. Backward taking min(self, right+1) // Missing last
What has to be changed for distance transform in 2D?
Replace left and right pixel above or left and below or right
What are the characteristics of parametric search in fitting? [4]
1. Parametric model with observed features
2. Infeasible to check all combinations of models
3. Membership criteria is not local
4. Hough transform (HT) with hough space (HS)
What is the idea of hough transform?
Transform edge pixels into lines and vote for them to deduce the best lines
How are elements of the image space saved in hough space? [3]
1. Save point (m,b) in HS for each line y=m*x+b // Undefined for vertical lines
2. Save line b=x*m+y in HS for each point (x,y)
3. Save (d,t) in accumulator array H with x*cos(t)-y*sin(t)=d // Polar representation
How can the elements in hough space be used? [2]
1. Line by p1 p2 in IS is crossing of two lines in HS
2. Vote for best line by more than two points
What is the hough transform algorithm? [4]
1. Init H(d,t) with 0
2. Forall edge pixel (x,y) and t=0 → 180 increase H(d,t)
3. Find maximizing (d,t)
4. Return line x*cos(t)+y*sin(t)=d
What are the extensions of the the hough transform algorithm? [4]
1. t=G in 2.
2. Give more vote for stronger edges
3. Change resolution of H
4. Change shape
What the characteristics of a circle used as shape in hough transform algorithm? [3]
1. (x-a)²+(y-b)²=r with fix r
2. For unkown r and G with cone (Unkown G) or line (Known G) in 3D
3. forall r t (Estimated with G or each): a=x-r*cos(t) and b=y+r*sin(t)
What are the notes for hough transform? [4]
1. Remove irrelevant pixels
2. Vote for neighbors as smoothing in H
3. Use direction of edge to reduce free parameters
4. Tag votes
What are the advantages of hough transform? [3]
1. Independent pixel process
2. Robustness to noise
3. Detection of multiple instances in a single pass
What are the disadvantages of hough transform? [3]
1. Fast increasing complexity
2. Not relevant peaks in parameter space
3. Picking good grid size
What are the characteristics of generalized hough transform? [2]
1. Detect arbitrary shape by boundary points P:={p1,…, pn} in IS
2. Compute displacement vectors r=a-pi and store in table indexed by G orientation t
