MULTI TENSOR INJECTION (MUTI) FOR ADJUSTING TEXTURE INTENSITY AND IMAGE COLOR IN FAST NEURAL STYLE TRANSFER

• Main Author: Hidayaturrahman
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/46434
• Institution: Institut Teknologi Bandung

When performing “fast” neural style transfers using the arbitrary style transfer approach, content images and style images used as input can be more varied than approaches that map the style of images into the model as used in real-time style transfers. In addition this approach is also able to provide results faster than the approach that performs direct optimization of the pixel value of the image as is done in the "slow" neural style transfer. Behind all these advantages, the arbitrary style transfer approach is only able to produce images with content composition and style in accordance with the loss function specified at the time of training. Therefore, an approach called Multi Tensor Injection (MuTI) was formulated. MuTI is able to control the color intensity and texture of images produced by the neural style transfer model without going through the process of retraining the model. In this study there are several scenarios used in conducting experiments in stages with the MuTI approach. Based on the results of experiments conducted, it was found that by using MuTI with the input in the form of normalized style image encoding features, it gives better results than other approaches. This can be concluded because this approach is able to suppress the value of content and style loss better than other approaches.