Publications

Understanding neural-network denoisers through an activation function perspective, Yuxiang Li, Bo Zhang and Raoul Florent

While neural network is able to denoise an image, the denoising mechanism behind is difficult to explain due to non-linear activation functions. Our paper focuses on visualizing learned patterns in each layer of a neural network while varying the activation function. We also introduce a trainable activation layer that converges to a soft-thresholding function which happens to be a widely used function in denoising methods.

Proc. IEEE International Conference on Image Processing (ICIP), 2017.

Fast de-streaking method using plain neural network, Yuxiang Li, Bo Zhang and Raoul Florent

Radon transform fundamentally underlies reconstructions from computed tomography. In this work, we propose a method that reduce the steaks on a reconstructed image: a de-streaker based on plain neural network. The objective is not to replace reconstruction methods, but to offer a fast post-processing that reduces artifacts from the output image, particularly when sinogram is unavailable.

Proc. IEEE International Conference on Image Processing (ICIP), 2017.

SHREC 2016: 3D Sketch-Based 3D Shape Retrieval, Bo Li et al.

In this workshop, we propose a non-learning method based on image classification. We transform 3D models to noised point cloud, then we use convolutional neural network to classify a 3D sketch in form of set of images. Our approach is the best among all non-learning methods. [website]

Proc. Eurographics 2016 Workshop on 3D Object Retrieval (3DOR), 2016.