I am trying to design a CNN that can do pixel wise segmentation of edges leaves in dense foliage agriculture images. Such as these:
On the basis of this article https://arxiv.org/pdf/1904.03124.pdf, two classes are defined, such as the external contours and the internal contours of the leaves boundaries. In addition, a multi-scale approach is used trough a unet like architecture and a auxiliary loss was used to learn the edge detection at different scales (which corresponds to the main idea of this article https://arxiv.org/pdf/1804.01646.pdf). I learn the network using a mIoU loss whose weight varies depending on the class and the scale. Finally, my last activation layer is a clipped ReLU. The results are starting to be good :
However, the network is not able to reconnect some internal edges. The following image shows a broken inner edge that does not continue to the next part (in blue) :
So I'm looking for some paper, git, codes, whatever that can improve the reconstruction of missing edges (inside or outside the CNN).