Object Segmentation with Neural Network Combined Grab Cut

Style transfer refers to the technique which applies the style of an artistic image to a real image which contains different contents than the artistic image. Nowadays, state-of-the-art results are obtained by using deep neural networks based style transfer methods. Recently, researches have been performed that apply the style of a single image to a whole video sequence. To give a feeling of a mixture of a real world and an animated world, we proposed a method that can apply the style of a single still image only on a selected object in the video sequence. In this paper, we propose an improved version of this method to obtain a more correct region of the object. The method combines the level set based segmentation and the GrabCut method together. The level set based segmentation suggests the foreground and the background colors to the Gaussian mixture model in the GrabCut method, and using this color suggestions, the GrabCut method cuts out the object region. Experimental results show that the proposed method can accurately select the object on which the object selective style transfer is applied.