Correção Automatizada de Perspectiva e Orientação de Imagens de Documentos

Abstract

Invoices and other high-value documents are usually printed on paper or in PDF format. Thus, automated document analysis is hampered by distortions in image digitization, which highlights the importance of document processing. This article addresses the challenges of automating document image processing and proposes an instance segmentation model with YOLOv8, perspective correction and background removal. A second processing stage includes noise filtering and orientation detection, using the Z-Score statistical method to improve accuracy in tilt correction. The orientation detection algorithm showed an average error of 0.108 degrees, while the document segmentation model showed an accuracy of 0.997. Therefore, the proposed algorithms can be applied as pre-processing in projects for the automated extraction of data from invoices and similar documents.