Ocenjevanje kakovosti slik poslovnih dokumentov z globokim učenjem na podlagi lokalnih regij

Krivec, Jan (author), Skočaj, Danijel (mentor)

Sistemi za optično prepoznavo znakov so občutljivi na kakovost vhodnih slik, nepravilna prepoznava pa pomeni napake v nadaljnji obdelavi. Obstoječi pristopi k ocenjevanju kakovosti dokument opišejo z eno samo oceno, ki ne pove niti kje na dokumentu so napake, niti ali so pomembni deli dokumenta okrnjeni. V delu predstavimo lokalizirano ocenjevanje kakovosti slik poslovnih dokumentov. Zgradimo sintetično množico 2299 računov, ki vsebuje oznake semantičnih regij dokumenta in v kateri je napaka strojnega branja izmerjena za vsako besedo posebej ter na njej naučimo enotni večopravilni model s skupnim kodirnikom, ki iz vhodne slike hkrati napove sliko intenzivnosti napake in segmentacijo pomenskih regij. Model na prej nevidenih predlogah dosega dokumentni SROCC 0,91, pri odločanju o sprejemljivosti dokumenta glede na pomembne regije pa preseže tudi teoretično zgornjo mejo vseh ocen na ravni celotne strani (AUC 0,92 proti 0,76). S poskusi na referenčni množici SmartDoc-QA pokažemo, da je tudi pri napovedi ene povprečne ocene kakovosti strani pomemben lokaliziran učni signal in ne samo arhitektura modela.

• Language: Slovenian
• Keywords: optična prepoznava znakov, strojno učenje, globoko učenje, globoke nevronske mreže
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-189099
• Publication date in RUL: 01.10.2026

Secondary language

• Language: English
• Title: Business document image quality assessment with deep learning based on local regions

Abstract

Optical character recognition systems are sensitive to the quality of input images and recognition errors propagate to downstream processes. Existing quality assessment approaches describe the document with a single score that reveals neither where the errors occur nor whether they affect its important parts. This thesis presents a localised quality assessment for business document images. We build a synthetic dataset of 2299 invoices, annotate the recognition error for every individual word and train a unified model with a shared encoder that predicts both the recognition error heatmap and a segmentation mask of semantic regions for the document. The model reaches a document-level SROCC of 0.91 on previously unseen templates and surpasses even the theoretical upper bound of any page-level score when classifying document acceptability based on semantically important regions (AUC 0.92 vs. 0.76). Experiments on the SmartDoc-QA dataset show that even approaches directed at page-level quality score benefit from a localised training signal.

• Keywords: optical character recognition, machine learning, deep learning, deep neural networks