Denoising and Deblurring Networks Quiz

Answers and explanations appear as you take the quiz.

1. What does DnCNN actually predict during training?
   • The noise residual, which is then subtracted from the input
   • The exact camera ISO setting
   • A text description of the scene
   • The depth map of the image
2. Why is blind deblurring considered harder than denoising?
   • Because images are always in color
   • Because the blur kernel describing how pixels smeared is usually unknown
   • Because deblurring needs a GPS signal
   • Because noise cannot be added synthetically
3. How are training pairs for these networks typically created?
   • By hand-drawing clean versions
   • By scanning film negatives
   • By adding synthetic noise or blur to clean photos
   • By recording audio alongside the image
4. What architecture is commonly used as the backbone for image restoration?
   • A simple linear regression
   • A decision tree
   • A word embedding lookup
   • A U-Net encoder-decoder
5. What is the key idea behind the Noise2Noise training approach?
   • It maps one noisy observation to another noisy one, needing no clean target
   • It requires perfectly clean target images
   • It only works on text
   • It uses no neural network at all
6. Why can adding perceptual or adversarial objectives be a tradeoff in restoration?
   • They make images impossible to display
   • They may improve visual naturalness but can favor plausible detail beyond pixel fidelity
   • They guarantee exact recovery of every pixel
   • They remove the need for training data
7. What is SIDD used for in image restoration research?
   • Measuring telescope mirror alignment
   • Ranking speech transcripts
   • Evaluating denoising on real smartphone camera noise
   • Testing particle-detector triggers
8. What is a risk when a generative restoration model sharpens an image?
   • It always reduces image dimensions
   • It converts every image to grayscale
   • It removes the need to inspect results
   • It may invent plausible detail that was not supported by the measurement