FASTER ART-CNN: AN EXTREMELY FAST STYLE TRANSFER NETWORK

Convolutional neural networks have recently found applications in the artistic realm, where style and content images can be combined to yield a new image. This paper presents a novel architecture for the performance of extremely high speed style transfer in the feed-forward mode, with a minimal qualitative decrease in image quality. This is accomplished by training a deconvolutional neural network to apply a specific style to the provided content image. To make the problem computationally tractable, the content images in this work are restricted to the dataset of Labeled Faces in the Wild. Our faster style transfer results favorably compare to the traditional backpropagation technique and another existing feed-forward technique. The real time performance of our Faster Art-CNN would be suitable for augmented reality, video conferencing and other computationally demanding applications.