Abstract
During ship drawing review, surveyors at classification societies manually compare original and revised drawings to verify that requested revisions have been reflected. This task is time-consuming and error-prone because composition inconsistency can disrupt one-to-one correspondence between drawing sets, requiring surveyors to match drawings before distinguishing subtle revisions from visual differences caused by alignment and quality inconsistencies in large-scale, information-dense drawings. This study proposed an automated method for revision identification using deep learning-based change detection. Corresponding drawings were matched using embedding-based image similarity to address composition inconsistency. To overcome the shortage of drawing data with labeled revision regions, synthetic training data were acquired from original drawings; synthetic revisions resembling actual revisions were generated through image inpainting and segmentation, while data augmentation reflected alignment and quality inconsistencies. A pretrained change detection model was fine-tuned on synthetic training data and applied to large, high-resolution drawings using an adapted sliding-window algorithm. Experiments on 40 drawing pairs from five shipyards and across different drawing types showed reliable, rapid matching. Revision identification achieved 98.13% precision, 98.99% recall, and a 98.56% F1-score, averaging 7.44 s per pair. The proposed method can help surveyors focus their review on identified revisions, thereby supporting efficient and consistent drawing review.
