Image Segmentation with a Deep Declarative Network

Lech Szymanski, Steven Mills, Garth Wales, Hayden McAlister

Abstract: Deep learning methods work extremely well for many image processing applications, but are hard to interpret, data hungry and prone to learning the wrong thing (if the training data is not well curated). Could we bootstrap the learning process with a classical algorithm that solves the problem analytically? In this work we test Deep Declarative Network (DDN), an implicit-differentiation-based framework, for embedding a classical method, such as Normalised Graph Cuts segmentation, into a neural network. It works, allowing us to train the entire system end-to-end. Unfortunately, the combined system is slow, memory intensive, and does not match the seg-mentation performance of the plain feed-forward neural network. Nevertheless, we demonstrate that DDN is a capable system for combining classical and neural-network based methods and that our proposed polarity-agnostic loss function is well suited for segmentation, and not merely segment classification.