Abstract
Surveyors of classification societies review shipyard drawings during the design stage to ensure safety and regulatory compliance. When revisions are required, the surveyors issue comments, and the designers modify and resubmit the drawings; this process is repeated until all requirements are satisfied. In practice, designers are expected to mark each revision with symbols such as revision clouds; however, these marks may be unintentionally omitted. Moreover, iterative revisions may lead to additional changes occurring beyond the marked areas, making the revision markings incomplete and less reliable. As a result, surveyors typically must rely on direct visual comparison of the original and revised drawings to verify the revisions, which is a significant challenge. This challenge is compounded by inconsistencies that arise across review cycles. Since designers submit revised drawings as a set in different ways—some omit unrevised drawings, others add new ones, and the ordering often differs—the original and revised sets rarely correspond one-to-one, creating a composition inconsistency that forces surveyors to first match each drawing to its counterpart. In addition, repeated scanning or exporting introduces alignment inconsistencies, such as shifts and rotations, and quality inconsistencies, such as differences in resolution and line width, that can mimic changes even when nothing was edited, making visual review even more difficult and error-prone.
Therefore, this study proposed a method for automatically identifying revisions in ship drawings using deep learning-based change detection technology that leverages surrounding context to detect revisions robustly. Corresponding drawings are first paired by encoding each one as a feature vector using an embedding-based image similarity approach and assigning globally optimal one-to-one matches with the Hungarian algorithm, resolving the composition inconsistency between the two sets before revision identification. Training such a change-detection model, however, requires a large number of original–revised drawing pairs with labeled revision regions, and such data are especially scarce in this field: ship drawings are rarely disclosed externally due to strict security requirements, so virtually no public datasets label the revised regions between drawing versions. To overcome this limitation, the proposed method introduced a technique that uses image inpainting and segmentation from the field of computer vision to generate synthetic revised drawings from original drawings, enabling the acquisition of training data without any real revised drawings. Specifically, we defined revisions as three representative types—addition, removal, and replacement—and generated synthetic revisions that closely resemble real revision patterns by reflecting these types. Removal and replacement were generated by a domain-fine-tuned inpainting model, whereas addition was handled by reinserting elements extracted via image segmentation into the same drawing, ensuring stylistic consistency. Together with augmentation that reflects real alignment and quality inconsistencies, this yielded, from the original drawings alone, 240,000 training triplets, each consisting of an original drawing, a synthetic revised drawing, and a mask labeling the revised regions.
A CNN-based change detection model was adopted for its strength in capturing fine, local changes in thin lines and symbols; the model, pretrained on a large-scale remote sensing dataset, was then fine-tuned on the synthetic data. Furthermore, to handle very large, high-resolution drawings, a sliding-window algorithm for change detection was applied at the inference stage. On 40 real drawing pairs from five shipyards and several drawing types, matching reached 100% accuracy in 37 seconds, and revisions were identified with a 98.56% F1-score (98.13% precision, 98.99% recall) in 7.44 seconds per pair. The near-99% recall shows that, in drawing review where missing revisions must be avoided, the change detection model trained on the synthetic dataset can rapidly and accurately identify revisions in actual drawings and present them to the surveyors. Consequently, the surveyors can perform rapid, intensive reviews centered on the identified areas, significantly enhancing the efficiency of the iterative drawing review process.
Therefore, this study proposed a method for automatically identifying revisions in ship drawings using deep learning-based change detection technology that leverages surrounding context to detect revisions robustly. Corresponding drawings are first paired by encoding each one as a feature vector using an embedding-based image similarity approach and assigning globally optimal one-to-one matches with the Hungarian algorithm, resolving the composition inconsistency between the two sets before revision identification. Training such a change-detection model, however, requires a large number of original–revised drawing pairs with labeled revision regions, and such data are especially scarce in this field: ship drawings are rarely disclosed externally due to strict security requirements, so virtually no public datasets label the revised regions between drawing versions. To overcome this limitation, the proposed method introduced a technique that uses image inpainting and segmentation from the field of computer vision to generate synthetic revised drawings from original drawings, enabling the acquisition of training data without any real revised drawings. Specifically, we defined revisions as three representative types—addition, removal, and replacement—and generated synthetic revisions that closely resemble real revision patterns by reflecting these types. Removal and replacement were generated by a domain-fine-tuned inpainting model, whereas addition was handled by reinserting elements extracted via image segmentation into the same drawing, ensuring stylistic consistency. Together with augmentation that reflects real alignment and quality inconsistencies, this yielded, from the original drawings alone, 240,000 training triplets, each consisting of an original drawing, a synthetic revised drawing, and a mask labeling the revised regions.
A CNN-based change detection model was adopted for its strength in capturing fine, local changes in thin lines and symbols; the model, pretrained on a large-scale remote sensing dataset, was then fine-tuned on the synthetic data. Furthermore, to handle very large, high-resolution drawings, a sliding-window algorithm for change detection was applied at the inference stage. On 40 real drawing pairs from five shipyards and several drawing types, matching reached 100% accuracy in 37 seconds, and revisions were identified with a 98.56% F1-score (98.13% precision, 98.99% recall) in 7.44 seconds per pair. The near-99% recall shows that, in drawing review where missing revisions must be avoided, the change detection model trained on the synthetic dataset can rapidly and accurately identify revisions in actual drawings and present them to the surveyors. Consequently, the surveyors can perform rapid, intensive reviews centered on the identified areas, significantly enhancing the efficiency of the iterative drawing review process.