Multiple Document Datasets Pre-training Improves Text Line Detection With Deep Neural Networks

Robust Text Line Detection in Historical Documents: Learning and Evaluation Methods

“…Their model provided second best results on the well-known cBAD dataset [26] while having a really light post-processing compared to the first best model. Lastly, in one of our previous works we presented Doc-UFCN [6], a FCN inspired by the previously cited systems and by the multi-modal model presented by Yang et al [27]. The encoder is composed of dilated convolutions while the decoder consists in transposed convolutions.…”

“…Their model follows an encoder-decoder architecture where the encoder is a ResNet-50 [28] pre-trained on natural scene images such as ImageNet [29]. Even if this method has shown good results on various tasks (page extraction, baseline detection or layout analysis), the inference time is still significant as shown in one of our early works [6]. Yang et al [27] also presented a model for multiple classes segmentation.…”

“…Object-level IoU P/R F1 P/R R@.85/.95 mAP@.65 mAP Barakat [23] Mechi [24] Renton [25] Doc-UFCN [6] dhSegment [7] Yang [27] Tarride [2] Soullard [30] Melnikov [31] To tackle this problem, metrics originally designed in the Information Retrieval community [34] have been adapted to images and used during the PASCAL VOC Challenge 2012 to compute the Precision at objectlevel. During this competition, the detection task was evaluated based on the Precision-Recall curve at objectlevel, where the detections are considered as true or false positives according to their area of overlap with the ground-truth objects.…”

“…All the datasets cited above have been used to train generic text line detection models using Doc-UFCN [6], dhSegment [7] and ARU-Net [8]. These models need pixel-level annotated images for training, we present in the next paragraphs the challenges faced to generate a unified annotated training set.…”

“…In this paper, we provide a fair and extensive evaluation of three state-of-the-art approaches for text line segmentation, Doc-UFCN [6], dhSegment [7] and ARU-Net [8] on a wide collection of historical datasets and using multiple metrics including a goal-directed one. We analyze the line segmentation metrics of the literature with regards to multiple datasets publicly available and show some label inconsistencies among the datasets.…”