Convolutional Autoencoder for Denoising Images of Active Nematics

The images produced in active nematics experiments are usually quite noisy. These noises (caused by the activities of materials, lighting, etc.) can significantly influence the downstream analyses and constraint our capability of understanding active nematics. Conventional denoising methods have limited power in improving the quality of those images. In this study, we developed a Deep Learning technique to tackle this problem. We designed a deep denoising model as a deep convolutional auto-encoder with skip-layers. The deep denoising model was trained using randomly chosen clean images of active nematics. We also designed a convolutional reconstruction method that uses our deep denoising model to remove noises in new active nematics images. Experimental results demonstrate the effectiveness of our approach.