• Kacper Kazaniecki

2018 Huawei AI Challenge - Image Denoising

Along with William Profit, we participated in the annual Huawei AI UK competition throughout November of 2018. The task was to create an image denoising algorithm. We were provided with a dataset of around 1000 4k image pairs of clean and noisy. They were split up into 3 equal categories: Foliage, Buildings and Text. After spending nearly two weeks in heavy research, we decided to base our method on the paper called "Image Restoration Using Convolutional Auto-encoders with Symmetric Skip Connections" by Mao et al. We augmented the data by flipping and rotating the images to giving us more data to train on. Since the original images were of 4k resolution, we decided to split them up into 256x256 squares. As one of the very few teams of 1st-year students competing against even some teams of PhD students (e.g. top 3 places were all PhD), we were really pleased to come 12th. We've learnt a lot during the competition and look forward to doing it next year.

24 views0 comments

Recent Posts

See All