AlioLeuchtmann/Qwen2.5_3B_VL_PDF_ROTATION_DETECTION_MK1
AlioLeuchtmann/Qwen2.5_3B_VL_PDF_ROTATION_DETECTION_MK1 is a 3 billion parameter vision-language model, fine-tuned from Qwen2.5-VL 3B, specifically for binary rotation detection of PDF page images. Trained on 12,000 annotated samples from 8GB of scanned PDF data, this model excels at identifying whether a document page is flipped or correctly oriented. Its primary use case is to enhance automated document processing pipelines by ensuring correct page orientation for downstream OCR and VLM systems, achieving 100% accuracy on 1992x1992 pixel images.
Loading preview...
Overview
AlioLeuchtmann/Qwen2.5_3B_VL_PDF_ROTATION_DETECTION_MK1 is a specialized vision-language model, fine-tuned from Qwen2.5-VL 3B, designed for robust binary classification of document page orientation. It determines if a PDF page is upside down (flipped by 180 degrees) or correctly oriented, responding with 'Yes' or 'No'. This model was trained on approximately 8GB of scanned PDF page image data, comprising 12,000 annotated samples, with a focus on document-specific rotation detection rather than general image classification.
Key Capabilities
- Highly Accurate Rotation Detection: Achieves 100% precision, recall, F1 score, and accuracy on 1992x1992 pixel images, significantly outperforming the base model.
- Document Processing Enhancement: Purpose-built to improve data quality and consistency in automated document workflows by correcting page orientation issues.
- Efficient Inference: Capable of processing multiple pages per second on consumer hardware like the RTX 4060 Ti, with optimal performance at 996x996 pixel input for speed and quality trade-off.
- Robustness: Demonstrates improved performance on higher-resolution inputs (1992x1992) even though trained on 996x996 images, and performs well on German test sets despite a mostly English training dataset.
Good for
- Automated Document Processing: Ideal for integration into pipelines where scanned PDF pages may be inadvertently oriented incorrectly.
- Improving OCR and VLM Performance: Ensures that downstream OCR systems and Vision-Language Models receive correctly oriented input, leading to more accurate information extraction.
- High-Volume Document Workflows: Its efficiency allows for rapid processing of large quantities of document pages, making it suitable for enterprise-level applications.