Quality Prediction for Large Scale HTR Confidence Is All You Need
Year: 2026
Type: conference-paper
Abstract: Quality prediction of handwritten text recognition (HTR) addresses the challenge of estimating HTR quality in production where no ground-truth data are available. Focusing on pipeline-based HTR (instance segmentation for regions/lines followed by line-level recognition), we formulate quality prediction as a supervised page-level regression task: given a single document image and the runtime outputs of the HTR pipeline, predict full-page quality metrics that reflect both line/region segmentation and transcription performance. We extract and investigate four families of features available in production: (i) handcrafted image features, (ii) learned document-image embeddings from a pretrained document transformer, (iii) text-based features from external resources, including lexicon matching and character n -gram language model coverage/perplexity, and (iv) model-native features derived directly from the pipeline outputs, specifically confidence scores. A key contribution is to move beyond mean confidence and instead model confidence distributions via histogram features computed over tokens, lines, and region/line segmentation outputs. We find that confidence distributions consistently improve quality prediction and, when combined with standard confidence features, largely subsume the information provided by image- and language-based features. This implies that an effective quality prediction system can be built using low-cost model-native features.
Authors: Erik Lenas, Viktoria Löfgren, Olof Karsvall
Institution: National Archives of Sweden
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $5,392
OpenAlex ID: W7204168937