Optimizing Small Target Detection in Engineering Drawings: A Scale-Aware Slicing Approach

The application of Optical Character Recognition (OCR) in digitizing engineering drawings faces challenges in handling large-scale drawings with densely packed textual information. Conventional OCR designs optimized for smaller documents result in distortions when applied to large-scale drawings, leading to missed detections. This paper focuses on enhancing OCR’s text detection capabilities for large-scale drawings, where conventional designs fall short. A novel sub-image slicing method is introduced to address small target detection in large-scale engineering drawings. The method divides the large-scale drawing into overlapping sub-images, processed by a text detection model, and aggregates results to enhance detection rates. A text scale-aware sub-image slicing approach dynamically adjusts sub-image sizes based on drawing dimensions, mitigating issues of missed detections. Experimental evaluations demonstrate the proposed method’s superiority over existing methods, addressing adaptability issues and scaling challenges. This work contributes significantly to advancing OCR technology, specifically in the context of large-scale engineering drawings.