Historical Handwritten Document Segmentation by Using a Weighted Loss

In this work we propose one deep architecture to identify text and not-text regions in historical handwritten documents. In particular we adopt the U-net architecture in combination with a suitable weighted loss function in order to put more emphasis on most critical areas. We define one weighted map to balance the pixel frequency among classes and to guide the training with local prior rules. In the experiments we evaluate the performance of the U-net architecture and of the weighted training on one benchmark dataset. We obtain good results using global metrics improving global and local classification scores.