Evidential fully convolutional network for semantic segmentation

Zheng Tong (1) , Philippe Xu (1) , Thierry Denoeux (1, 2)

1 Heudiasyc - Heuristique et Diagnostic des Systèmes Complexes [Compiègne]
2 IUF - Institut universitaire de France

We propose a hybrid architecture composed of a fully convolutional network (FCN) and a Dempster-Shafer layer for image semantic segmentation. In the so-called evidential FCN (E-FCN), an encoder-decoder architecture first extracts pixel-wise feature maps from an input image. A Dempster-Shafer layer then computes mass functions at each pixel location based on distances to prototypes. Finally, a utility layer performs semantic segmentation from mass functions and allows for imprecise classification of ambiguous pixels and outliers. We propose an end-to-end learning strategy for jointly updating the network parameters, which can make use of soft (imprecise) labels. Experiments using three databases (Pascal VOC 2011, MIT-scene Parsing and SIFT Flow) show that the proposed combination improves the accuracy and calibration of semantic segmentation by assigning confusing pixels to multi-class sets.

Mots clés: Evidence theory; belief function; fully convolutional network; decision analysis; semantic segmentation

Domaines: Intelligence artificielle [cs.AI]

Zheng Tong, Philippe Xu, Thierry Denoeux. Evidential fully convolutional network for semantic segmentation. Applied Intelligence, 2021, 51 (9), pp.6376-6399. ⟨10.1007/s10489-021-02327-0⟩. ⟨hal-03511108⟩