CIFAR-10 dataset consisting of:
- 60,000 color images (50,000 training and 10,000 testing)
- Each of size 32 x 32 (with 3 RGB channels)
- 10 classes with 6,000 images per class
Auto-Encoders
- Convolutional and Transposed Convolutional blocks have been used for the process of encoding and decoding respectively
- The encoding, as well as the decoding blocks, consists of Convolutional Layer, Batch Normalization and ReLU activation
- For encoding, 4 convolutional blocks with downsampling (using stride=2) as well as 1 convolutional block without downsampling that encodes an input image of size (32, 32, 3) to size (2, 2, 256)
- For decoding, 4 deconvolutional blocks with upsampling (using stride=2) and interleaving concatenations (concatenating 1-3, 2-2, ,3-1) as well as 1 final deconvolutional block to decode (or reconstruct) input of size (2, 2, 256) to size (32, 32, 3)
Noise level 1 – Normal distribution Noise with σ = 0.1 and μ=0.0 added on original image (0 to 1 scale)
Noise level 2 – Normal distribution Noise with σ = 0.3 and μ=0.0 added on original image (0 to 1 scale)
Noise level 3 – Normal distribution Noise with σ = 0.2 and μ=0.0 added on original image (0 to 1 scale)
The evaluation metric for Denoising used is Mean Squared Error (MSE) and the results for different noise levels are tabulated below:
| Noise Level | MSE due to Noise | MSE after Denoising | % MSE Reduction |
|---|---|---|---|
| Noise Level-1 | 0.0093 | 0.0018 | 80.5455 % |
| Noise Level-2 | 0.0642 | 0.0053 | 91.6092 % |
| Noise Level-3 | 0.0628 | 0.0037 | 94.0540 % |
Results for noise level 1:
Results for noise level 2:
Results for noise level 3:
