The paper, A. Krull, et al., "Noise2Void - Learning Denoising From Single Noisy Images," 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)., proposed a "blind-spot" kernel for image denoising, where the "blind-spot" kernel means a receoptive field excludes the center pixel. This prevents a neural network to learn the identity in a self-supervised training scheme.
The original paper mentioned that implementing a "blind-spot" kernel is not trivial, so instead the authors used a so called masking scheme (details can be found in the paper). However, to the best of my knowledge, this kernel can be approximated using a kernel constraint that sets the center weight to be 0. Note that implementing the blind-spot as a kernel constraint is only an approximation, as a constraint is applied after a kernel was updated by training. This differs from that in the original paper.
An example usage in the Conv2D layer in Keras:
x = Conv2D(nf, 3, activation='relu', padding='same', kernel_constraint=blind_spot(filt_sz=3))(x)
Here is an example result obtained using a residual-ConvNet-based blind spot network:
