Overview
Noise2Void (N2V) is a self-supervised deep learning method for image denoising. It operates on single noisy images, eliminating the need for clean target images or noisy image pairs, enabling its use in scenarios where acquiring training targets is impractical, especially in biomedical imaging. The architecture leverages a modified UNet to reduce checkerboard artifacts, replacing MaxPool layers with BlurPool layers, rolling back the residual-UNet to a non-residual UNet, and eliminating skip connections at the uppermost UNet level. N2V's training scheme involves masking or creating blind spots in the input image and predicting the masked pixel intensities from neighboring pixels. Although N2V cannot outperform methods with more information, its performance compares favorably to training-free denoising approaches. The implementation is based on TensorFlow and has been tested with Python 3.9 and TensorFlow versions 2.7, 2.10, and 2.13.
Common tasks
