Fast Convergence of Deep Learning Model on Image Inpainting Using Wavelet Activation Functions
碩士 === 輔仁大學 === 數學系碩士班 === 107 === The neural network may not be novel. As early in the 19th century, Warren McCulloch and Walter Pitts et al. [1] proposed a neural-framework and an algorithm called threshold logic to create a neural computational model of the network. But it was silent because the...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | zh-TW |
| Published: |
2019
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/duzhep |
| Summary: | 碩士 === 輔仁大學 === 數學系碩士班 === 107 === The neural network may not be novel. As early in the 19th century, Warren McCulloch and Walter Pitts et al. [1] proposed a neural-framework and an algorithm called threshold logic to create a neural computational model of the network. But it was silent because the algorithm was not mature at the time and the computer system did not have enough capacity to handle the expansive computational time for such large neural networks. In the current, the neural network algorithm has come back in the name of deep learning due to the back propagation method , various algorithms, and the computational power. This paper proposes that using the wavelet function as an activation function in the neural network. The method is more accurate and faster than the results using traditional activation functions. We choose the Deep Convolutional Generative Adversarial Networks to inpaint the missed image as the subject of experiments. Furthermore to explore the extent of the repaired images that can be shown with different missing proportions of images.
|
|---|