Pretrained models are available here
One of the fundamental challenges in the field of image processing and computer vision is image denoising , where the underlying goal is to estimate the original image by suppressing noise from a noise-contaminated version of the image.It is a well-studied problem in Computer Vision.
For this task a fully Convolutional auto-encoder network was used. The network consists of multiple convolution and transposed convolution(deconvolution) layers, learning mapping from corrupted images to original images.
- The convolutional layers capture the contents of the image while eliminating the noise.
- The transposed convolution layers upsample the feature maps and *recover the details of the image.
Skip connections are added from a convolutional layer to its corresponding mirrored transposed convolution layer. These skip connections exhibit two major advantages:
- They pass image details from convolutional layers to transposed convolution layers which is useful in recovering the clean image.
- They allow signal to be back-propagated directly to the layers at the starting and hence tackling the problem of vanishing gradients.
- The network is fully convolutional and transposed convolution(deconvolution).
- The convolutional and transposed convolution layers are symmetric.
- ReLUs used after every convolutional and transposed convolution layer.
- Pooling/unpooling is not used as pooling discards useful image details that are useful for these kind of tasks.
- Skip connections are added from a convolutional layes to its corresponding transposed convolution layer.
- CIFAR-10 dataset is used in this task.
- An additive Gaussian noise with zero mean and standard deviation of 0.1, 0.3, 0.5, 0.7, 1.0 is used in this task.
- Peak signal-to-noise ratio(PSNR) is used to compare the performance. It is given by
psnr = 20 * np.log10(max_pixel / np.sqrt(mse))
mse- mean square error; max_pixel- Max value of pixel in an image
- Network performance on different values of standard deviation:
Images showing original images(row 1), noisy images(row2), denoised images(row 3) for standard deviation 0.1
More results are shown in the results.pdf