Comparative analysis of the denoising effect of unstructured vs. convolutional autoencoders
Sajid Majeed, Yusra Mansoor, Sana Qabil, Farooq Majeed, Behraj Khan · 2020
Adding noise to data affects the prediction of a discriminator. Some like a deep neural network extract features directly from inputs, where the quality of the features may be affected by the amount of the noise. A deep neural model specifically meant for feature extraction is the autoencoder, while it has also been extended to perform denoising. In this paper, we investigate the denoising effect of an encoder on different nature as well as different amounts of additive noise. The experiments are evaluated on a linearized autoencoder as well as a convolutional autoencoder, which is especially meant for image data. Denoising Autoencoders (DAE) and Convolutional Denoising Autoencoders (CDAE) are evaluated by introducing with Gaussian, Salt and Pepper, and Poisson types of noise with a factor of 0.5. The results shows 0.12, 0.09, 0.47 Mean Squared Error (MSE) for DAE and 0.13, 0.10 and 0.9 MSE in case of CDAE with the same amount of noise factor added, alluding to the insight that a lack of focus on structure in the model may help it focus more on the denoising task.