preview

Convolutional Networking Essay

Better Essays

2.1 Styling using Convolutional Neural Networks
The initial work on style transferring using convolutional neural networks was brought forth by Leon Gatys, Alexander Ecker and Matthias Bethge [1] in which style representations were extracted from images. This involved superimposing the style image onto the content image such that the semantic details of the content image were not lost as shown in Figure 1 [1].

Figure 1: Example of using the neural style transfer method advocated by Gatys et al to transfer the style of the reference photo onto the original photo such that the stylized photo contains the semantic content of the original photo and the style of the reference photo

The paper provides a detailed explanation of how the …show more content…

2.2. Neural Style Transfer
Becattini et al [2] provide a thorough survey of neural style transfer literature. It highlights the contemporary problems of the existing models and also discusses future scope. It also presents various evaluation techniques to contrast outputs obtained through different neural style transfer methods. A broad classification of neural methods into descriptive and generative neural methods where the former is used to transfer styles by updating pixel information in the images whereas the latter optimizes a model and generates an image with varying styles in a single pass. Figure 3 provides a visually contrasting stylized image when a content image is stylized using both the mentioned methods.

Figure 3: (Starting from the left) a) content image b) style image c) Descriptive method results with brush size control d) Generative method without brush size control

The paper improves upon the existing methods mentioned above while being intent on process architecture preservation but improves on the performances by tweaking environmental parameters such as loss function modification, emphasizing on spatial arrangements to style details. Extensions to existing transfer methods which were developed for simple images were discussed, this included style transferring for doodles, head portraits, single user specified objects and video frames. The parameters mentioned in the paper were also broadly classified

Get Access