Has there been work in deep learning where filters are normalized with priors similar to 'box blurs' or other sorts of traditional CV filters, rather than being initialized to random continuous numbers at start of training?

A convolution multiplies a filter element-wise with the corresponding pixels where the filter is located and then sums these values. Then the filter moves across and down the image. In this example, the filter has all the same values. It therefore does an average of all the pixels surrounding each pixel, which blurs the image.