Mask Convolution for Filtering on Irregular-Shaped Image

In this paper, we proposed the concept of mask convolution, which operates on irregular shaped foreground, and could gain improvement on the margin of the object. In particular, we prove that this operation could be done in an efficient way. The extra calculation than normal convolution only lies on mask image convolution, and one time elements-wise multiply and division. To verify the necessity of mask convolution, we use mean filtering on retina images to eliminate spots. Compared to normal convolution, the proposed method could obtain clearer image on the margin of foreground, which is beneficial for subsequent CNN-based vessel segmentation model. We also proved that this method can be extended to other convolution kernels such as Gaussian and sharpen filters.