Spatialising uncertainty in image segmentation using weakly supervised convolutional neural networks: A case study from historical map processing

Convolutional neural networks (CNNs) such as encoder-decoder CNNs have increasingly been employed for semantic image segmentation at the pixel-level requiring pixel-level training labels, which are rarely available in real-world scenarios. In practice, weakly annotated training data at the image patch level are often used for pixel-level segmentation tasks, requiring further processing to obtain accurate results, mainly because the translation invariance of the CNN-based inference can turn into an impeding property leading to segmentation results of coarser spatial granularity compared with the original image. However, the inherent uncertainty in the segmented image and its relationships to translation invariance, CNN architecture, and classification scheme has never been analysed from an explicitly spatial perspective. Therefore, the authors propose measures to spatially visualise and assess class decision confidence based on spatially dense CNN predictions, resulting in continuous decision confidence surfaces. They find that such a visual-analytical method contributes to a better understanding of the spatial variability of class score confidence derived from weakly supervised CNN-based classifiers. They exemplify this approach by incorporating decision confidence surfaces into a processing chain for the extraction of human settlement features from historical map documents based on weakly annotated training data using different CNN architectures and classification schemes.