Fast and accurate page orientation detector for document analysis systems

Accurate detection of page orientation is a crucial preprocessing step for reliable document layout analysis (DLA) and optical character recognition (OCR). However, most existing systems treat orientation detection as a secondary task, resulting in reduced robustness when processing rotated documents. In this paper, we introduce a novel vertical orientation detector that combines efficient text chunk analysis with an iterative aggregation strategy, powered by a lightweight convolutional neural network. Our approach is extensively evaluated on a representative subset of the DocBank dataset, where it achieves high accuracy on both upright and rotated pages and demonstrates consistent robustness across diverse layouts. Compared to the state-of-the-art open-source OCR frameworks such as PaddleOCR and Tesseract OCR, our solution delivers superior accuracy on rotated documents, faster processing speeds, and lower memory consumption. These results highlight the potential of our method for integration into high-throughput, automated, and resource-constrained document processing pipelines.