Treffer: Image manipulation using convolutional neural network

M.Phil. ; A variety of low-level vision tasks are ill-posed, since multiple output images correspond to the same input image. These ill-posed tasks include super-resolution, denoising, deblurring, matting, inpainting etc. Previous methods usually require assumptions or priors to solve these ill-posed problems. When the assumptions does not hold, the algorithm performance often degenerates much. For example, uniform motion deblurring algorithms cannot generalize well to dynamic scenes due to the uniform blur kernel assumption is no longer satisfied. However, with the recent progress in convolutional neural networks (CNNs), these ill-posed problems benefit much from effective network structures and large volumes of paired training data. In this thesis, we aim at solving two low-level vision problems, i.e., dynamic scene deblurring and portrait image matting, by introducing new network structures and building new training and evaluation dataset for respective tasks. ; In the first part, we analyze parameter strategies for the deblurring networks i.e., parameter independence scheme in [36] and the parameter sharing scheme in [55], and propose a new selective sharing scheme with independent and shared modules. Inside the subnetwork in each scale, we propose a new nested skip connection structure for the nonlinear transformation modules to replace stacked convolution layers or residual blocks. Besides, we build a new large dataset of blurred/sharp image pairs towards better restoration quality. Comprehensive experimental results show that the parameter selective sharing scheme, nested skip connection structure, and the new dataset all significantly improve performance to set a new state-of-the-art in dynamic scene deblurring. ; In the second part, we propose an automatic portrait image matting system. This method does not need any user interaction, which was however essential in most previous approaches. In order to accomplish this goal, a new end-to-end CNN based framework is proposed to take the input of a portrait ...