Analyzing the Impact of Introducing Convolutional Layers in Variational Autoencoders

Sammanfattning

Variational autoencoders (VAEs) are probabilistic models which can be used for both image denoising and image generation through latent space sampling. Convolution neutral networks (CNNs) are widely used, although to what extent convolutional layers positively affect VAEs is not always clear. Therefore, this thesis implements and compares two non-convolutional and two convolutional VAEs to quantitatively investigate some of the effects of introducing convolutional layers into VAEs for image denoising and image generation. Manual and grid search methodologies were employed to find suitable models to be used in the comparisons. The scope of this thesis is limited to small networks and a dataset containing very small homogeneous images. The results from the comparison did not show significant differences in denoising performance between the non-convolutional and convolutional models. The differences between the models were less than 0.2% across both the mean structural similarity index measure and the peak signal-to-noise ratio metrics. Furthermore, the results showed that the non-convolutional model vastly outperformed the convolutional model in image generation, with the non-convolutional scoring three times better than the convolutional according to the Fréchet inception distance metric. On the contrary, convolutional layers did seem to show benefits through a more numerically stable training and they also allowed for deeper networks while improving convergence. These benefits were especially notable in denoising task, where the convolutional model converged much faster than its counterpart. In summary, introducing of convolutional layers did not invariably enhance the performance of VAEs in neither image denoising nor image generation tasks, indicating that convolutional layers should be used with some consideration. However, they do seem to open possibilities for deeper and more stable models that could be capable of surpassing the boundaries of non-convolutional models for more complex tasks.