A deep-learning team is augmenting a dataset of normalized (0 - 1) printed-circuit-board images. Besides rotations and flips, they want to add spurious noise so the classifier stays robust to sensor artifacts without changing the images' average brightness. Which parameter choice best meets these goals?
Add uniform noise drawn from the interval [0.0, 0.5] to every pixel
Randomly set 50 % of the pixels to zero to mimic dropout
Add pixel-wise Gaussian noise with mean = 0.3 and standard deviation = 0.05
Add pixel-wise Gaussian noise with mean = 0 and standard deviation = 0.05
Injecting zero-mean Gaussian noise perturbs individual pixels in a way that mimics real sensor noise while keeping the expected pixel value unchanged, so overall image brightness is preserved. A modest standard deviation such as 0.05 (on a 0 - 1 scale) is large enough to discourage the network from memorizing exact pixel values but small enough to avoid obliterating salient features.
The option that adds Gaussian noise with mean = 0 and σ = 0.05 therefore embodies sound spurious-noise augmentation practice.
Adding noise with a positive mean (0.3) systematically brightens images, introducing bias rather than pure noise. Uniform noise in the 0 - 0.5 range also shifts brightness upward on average. Replacing half the pixels with zeros is an aggressive cutout/occlusion strategy, not typical spurious noise, and would severely distort image content.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What is Gaussian noise, and why is it used in data augmentation?
Open an interactive chat with Bash
Why is it important to preserve the average brightness of images when augmenting datasets?
Open an interactive chat with Bash
What is the difference between spurious noise and aggressive transformations like dropout?