1. Which of the following is the primary goal of Super Resolution in image processing?
(A) Image segmentation
(B) Noise reduction
(C) Enhancing image resolution
(D) Image compression
2. What type of learning does SRCNN use to perform super-resolution?
(A) Reinforcement learning
(B) Supervised learning
(C) Unsupervised learning
(D) Self-supervised learning
3. What is the role of convolutional layers in SRCNN?
(A) To compress the image
(B) To extract and learn features
(C) To label image pixels
(D) To reduce image resolution
4. Which of the following is a limitation of SRCNN?
(A) Uses GAN-based architecture
(B) Cannot work on grayscale images
(C) Limited to shallow architecture
(D) Requires manual feature extraction
5. Which model introduced adversarial loss for super resolution?
(A) SRCNN
(B) VDSR
(C) ESRGAN
(D) Bicubic Interpolation
6. What is the function of the discriminator in ESRGAN?
(A) Downsamples the image
(B) Distinguishes real and fake high-resolution images
(C) Segments the image
(D) Encodes the image into latent space
7. Which of the following interpolation techniques is commonly used for initial upscaling in SRCNN?
(A) Bilinear
(B) Nearest Neighbor
(C) Bicubic
(D) Gaussian
8. What does the ‘Residual-in-Residual Dense Block’ (RRDB) improve in ESRGAN?
(A) Image compression
(B) Noise robustness
(C) Gradient flow and performance
(D) Feature normalization
9. What kind of loss is used in ESRGAN to ensure perceptual quality?
(A) Binary cross-entropy loss
(B) Mean squared error
(C) Perceptual and adversarial loss
(D) Categorical loss
10. Which dataset is commonly used to train super-resolution networks?
(A) COCO
(B) ImageNet
(C) DIV2K
(D) Cityscapes
11. Which layer in SRCNN is responsible for reconstructing the high-resolution image?
(A) First convolution layer
(B) Feature mapping layer
(C) Non-linear mapping layer
(D) Reconstruction layer
12. What is the effect of using a deeper network in super-resolution tasks?
(A) Decreases training time
(B) Reduces overfitting
(C) Improves feature learning and image quality
(D) Limits the resolution
13. Which performance metric is typically used to evaluate super-resolution results?
(A) IoU
(B) PSNR
(C) AUC
(D) ROC
14. Which GAN-based architecture is designed specifically for photo-realistic image enhancement?
(A) DCGAN
(B) CycleGAN
(C) SRGAN
(D) StyleGAN
15. What distinguishes ESRGAN from SRGAN?
(A) Use of deeper convolution layers
(B) Inclusion of RRDB blocks and improved perceptual loss
(C) Focus on image segmentation
(D) It uses autoencoders
16. In the context of super resolution, what does PSNR stand for?
(A) Pixel Sensitivity and Noise Ratio
(B) Peak Signal-to-Noise Ratio
(C) Pre-Scaled Noise Reduction
(D) Pixel Strength Normalization Ratio
17. Why is perceptual loss important in ESRGAN?
(A) Improves compression ratio
(B) Enhances pixel-wise accuracy
(C) Maintains visual quality as perceived by humans
(D) Increases training speed
18. Which layer operation is avoided in ESRGAN to preserve texture?
(A) Downsampling
(B) Upsampling
(C) Batch normalization
(D) Max pooling
19. What is the primary drawback of using only MSE loss in super-resolution tasks?
(A) Produces very blurry outputs
(B) Results in noisy reconstructions
(C) Requires too much training data
(D) Cannot optimize PSNR
20. Which model uses perceptual loss calculated using VGG features?
(A) SRCNN
(B) FSRCNN
(C) ESRGAN
(D) LapSRN
21. What is the input to the SRCNN model?
(A) Low-resolution image
(B) High-resolution image
(C) Image difference
(D) Latent feature map
22. What does the term ‘deep’ in deep learning refer to in super-resolution networks?
(A) Use of multiple layers
(B) Training on large datasets
(C) Fast inference
(D) Image depth information
23. Why is bicubic interpolation often used before feeding images to SRCNN?
(A) It improves memory usage
(B) It increases contrast
(C) It upsamples the image to desired size
(D) It reduces noise
24. Which component of GAN learns to generate images?
(A) Discriminator
(B) Reconstructor
(C) Generator
(D) Autoencoder
25. What is the purpose of using skip connections in super-resolution networks?
(A) For image downscaling
(B) For better color enhancement
(C) To preserve low-level features and gradients
(D) To reduce computation
26. Which domain are CNN-based super-resolution models primarily applied in?
(A) Frequency domain
(B) Spatial domain
(C) Wavelet domain
(D) Feature domain
27. Which of the following super-resolution models is the most basic CNN-based method?
(A) ESRGAN
(B) SRGAN
(C) SRCNN
(D) FSRCNN
28. What does FSRCNN stand for?
(A) Fully Supervised Recurrent Convolutional Neural Network
(B) Fast Super-Resolution Convolutional Neural Network
(C) Feature Scalable Residual Convolutional Network
(D) Filtered Signal Resolution Convolution Network
29. Why is FSRCNN faster than SRCNN?
(A) Uses smaller images
(B) Works directly on low-resolution images
(C) Replaces convolution with pooling
(D) Requires fewer training samples
30. What advantage does ESRGAN have over traditional upsampling techniques?
(A) Higher compression
(B) Realistic texture generation
(C) Faster inference
(D) Reduces model size
31. Which layer in ESRGAN helps to avoid vanishing gradients?
(A) RRDB
(B) Dropout
(C) Pooling
(D) Sigmoid
32. What is a typical size for training patches in super-resolution networks?
(A) 8×8
(B) 16×16
(C) 32×32
(D) 64×64
33. Which of the following is not an evaluation metric for super-resolution?
(A) PSNR
(B) SSIM
(C) BLEU
(D) LPIPS
34. What is SSIM used to measure in super-resolution results?
(A) Image size
(B) Training speed
(C) Structural similarity between images
(D) Color histograms
35. Which technique helps avoid checkerboard artifacts during upsampling?
(A) Transposed convolution
(B) Nearest neighbor upsampling + convolution
(C) Dropout
(D) Batch normalization
36. Why is VGG network used in perceptual loss?
(A) It is lightweight
(B) It helps reconstruct low-resolution images
(C) It extracts high-level semantic features
(D) It reduces training time
37. Which of the following is true about adversarial loss in ESRGAN?
(A) It increases pixel-wise accuracy
(B) It ensures natural image textures
(C) It trains only the discriminator
(D) It performs image segmentation
38. What does the term “hallucination” refer to in image super-resolution?
(A) Noise filtering
(B) Compression artifacts
(C) Realistic texture synthesis
(D) Data augmentation
39. What is a challenge of using GANs in super-resolution?
(A) Too many pooling layers
(B) Image sharpness
(C) Instability during training
(D) Fixed receptive fields
40. What is the main contribution of SRGAN?
(A) High PSNR values
(B) Realistic high-frequency details
(C) Large model size
(D) Reduced computation
41. Which method is most suitable for real-time applications among these?
(A) SRCNN
(B) ESRGAN
(C) FSRCNN
(D) SRGAN
42. Why do deeper networks tend to perform better in super-resolution?
(A) Reduced model size
(B) Better feature abstraction
(C) Lower memory usage
(D) Fewer parameters
43. Which technique helps ESRGAN generate sharper textures?
(A) Batch normalization
(B) L1 loss
(C) Adversarial loss
(D) ReLU activation
44. What is the main reason for using residual learning in super-resolution networks?
(A) Faster input processing
(B) Enables better low-level feature reuse
(C) Reduces overfitting
(D) Increases dataset size
45. Which method was the first to introduce deep learning for super resolution?
(A) ESRGAN
(B) FSRCNN
(C) SRCNN
(D) SRGAN
46. Which model uses pixel shuffle operation for upsampling?
(A) SRCNN
(B) SRGAN
(C) ESPCN
(D) ESRGAN
47. What is LPIPS used to evaluate?
(A) Loss gradient
(B) Dataset similarity
(C) Learned perceptual similarity
(D) Latent parameter statistics
48. Which type of noise affects high-resolution image outputs most significantly?
(A) Salt-and-pepper noise
(B) Compression noise
(C) Blur artifacts
(D) Gaussian noise
49. What is the primary input of a discriminator in ESRGAN?
(A) Blurry image
(B) Real or generated high-resolution image
(C) Latent vectors
(D) Downsampled patches
50. Which component of ESRGAN primarily contributes to generating fine image details and textures?
(A) Reconstruction layer
(B) RRDB (Residual-in-Residual Dense Block)
(C) Pooling layer
(D) Batch normalization
