Automatic document digitization is now a central challenge in many application contexts, since archives, forms, and texts acquired as images often exhibit highly heterogeneous characteristics. Traditional OCR systems perform well on regular printed documents, but show clear limitations when the content is handwritten, degraded, or composed of mixed portions. In the current state of the art, many approaches are effective only in specific scenarios, without ensuring truly adaptive management of the different textual typologies present within the same image. This work proposes an adaptive OCR pipeline for intelligent document digitization, designed to dynamically select the most suitable recognition engine according to the nature of the text. The innovative element of the system lies in the integration of a visual pre-analysis stage, a convolutional patch-based classifier to distinguish printed text, handwritten text, and mixed-content documents, and two specialized OCR modules, respectively dedicated to the recognition of printed text and more complex text. The process is completed by a linguistic post-processing stage aimed at correcting residual errors and improving the quality of the final output. The results obtained show that the proposed approach increases robustness, accuracy, and flexibility compared with non-adaptive solutions, confirming the effectiveness of the adopted strategy in handling heterogeneous and realistic documents.
La digitalizzazione automatica dei documenti rappresenta oggi una sfida centrale in molti contesti applicativi, poiché archivi, moduli e testi acquisiti in formato imma gine presentano spesso caratteristiche molto eterogenee. I sistemi OCR tradizionali offrono buone prestazioni sui documenti stampati regolari, ma mostrano limiti evi denti quando il contenuto è manoscritto, degradato o composto da porzioni miste. Allo stato dell’arte, molti approcci risultano infatti efficaci solo in scenari specifi ci, senza garantire una gestione realmente adattiva delle diverse tipologie testuali presenti nella stessa immagine. Questo lavoro propone una pipeline OCR adattiva per la digitalizzazione intel ligente dei documenti, progettata per selezionare dinamicamente il motore di rico noscimento più adatto in funzione della natura del testo. L’elemento innovativo del sistema risiede nell’integrazione tra una fase di pre-analisi visiva, un classifica tore basato su patch convoluzionali per distinguere testo stampato, manoscritto e contenuti misti, e due moduli OCR specializzati, rispettivamente dedicati al rico noscimento del testo stampato e di quello più complesso. Il processo è completato da una fase di post-processing linguistico finalizzata a correggere errori residui e migliorare la qualità dell’output finale. I risultati ottenuti mostrano che l’approccio proposto consente di aumentare robustezza, accuratezza e flessibilità rispetto a so luzioni non adattive, confermando l’efficacia della strategia adottata nella gestione di documenti eterogenei e realistici.
VERSO UNA DIGITALIZZAZIONE INTELLIGENTE DEI DOCUMENTI: ROUTING DINAMICO DI MODELLI OCR E INTEGRAZIONE DI LARGE LANGUAGE MODELS
CASSANO, FRANCESCO
2024/2025
Abstract
Automatic document digitization is now a central challenge in many application contexts, since archives, forms, and texts acquired as images often exhibit highly heterogeneous characteristics. Traditional OCR systems perform well on regular printed documents, but show clear limitations when the content is handwritten, degraded, or composed of mixed portions. In the current state of the art, many approaches are effective only in specific scenarios, without ensuring truly adaptive management of the different textual typologies present within the same image. This work proposes an adaptive OCR pipeline for intelligent document digitization, designed to dynamically select the most suitable recognition engine according to the nature of the text. The innovative element of the system lies in the integration of a visual pre-analysis stage, a convolutional patch-based classifier to distinguish printed text, handwritten text, and mixed-content documents, and two specialized OCR modules, respectively dedicated to the recognition of printed text and more complex text. The process is completed by a linguistic post-processing stage aimed at correcting residual errors and improving the quality of the final output. The results obtained show that the proposed approach increases robustness, accuracy, and flexibility compared with non-adaptive solutions, confirming the effectiveness of the adopted strategy in handling heterogeneous and realistic documents.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14253/109501