Adaptable Deep Learning and Probabilistic Graphical Model System for Semantic Segmentation

Abstract

Semantic segmentation algorithms based on deep learning architectures have been applied to a diverse set of problems. Consequently, new methodologies have emerged to push the state-of-the-art in this field forward, and the need for powerful user-friendly software increased significantly. The combination of Conditional Random Fields (CRFs) and Convolutional Neural Networks (CNNs) boosted the results of pixel-level classification predictions. Recent work using a fully integrated CRF-RNN layer have shown strong advantages in segmentation benchmarks over the base models. Despite this success, the rigidity of these frameworks prevents mass adaptability for complex scientific datasets and presents challenges in optimally scaling these models. In this work, we introduce a new encoder-decoder system that overcomes both these issues. We adapt multiple CNNs as encoders, allowing for the definition of multiple function parameter arguments to structure the models according to the targeted datasets and scientific problem. We leverage the flexibility of the U-Net architecture to act as a scalable decoder. The CRF-RNN layer is integrated into the decoder as an optional final layer, keeping the entire system fully compatible with back-propagation. To evaluate the performance of our implementation, we performed experiments on the Oxford-IIIT Pet Dataset and to experimental scientific data acquired via micro-computed tomography (microCT), revealing the adaptability of this framework and the performance benefits from a fully end-to-end CNN-CRF system on a both experimental and benchmark datasets.