Min-Chul Kong, Myung-Il Roh, In-Chang Yeo, Ki-Su Kim, Jeongyoul Lee, Jongoh Kim, Gapheon Lee, "A Detection Method of Objects with Text in Drawings Based on Deep Learning", Proceedings of ISOPE(International Society of Offshore and Polar Engineers) 2023, Ottawa, Canada, 2023.06.19-23
International Conference
2023.02.13 12:23
Min-Chul Kong, Myung-Il Roh, In-Chang Yeo, Ki-Su Kim, Jeongyoul Lee, Jongoh Kim, Gapheon Lee, "A Detection Method of Objects with Text in Drawings Based on Deep Learning", Proceedings of ISOPE 2023, Ottawa, Canada, 2023.06.19-23
조회 수 134
| 첨부 '1' |
|---|
| Abstract | Drawings such as piping and instrument diagrams (P&ID) or ship safety plans have various equipment and components (lines, signs, and text). Every drawing has rules to express these objects with specific symbols. Generally, these drawings are very complex because they are large in size and consist of relationships between several objects. Therefore, the drawing investigators spend countless time and labor. For the above reasons, this study uses a deep learning model that has been actively researched recently. An object detection model based on deep learning can quickly find various objects within the drawing. However, the drawing differs from the common images in size and characteristics, generally used as an input in deep learning. Therefore, we proposed a series of procedures for applying the deep learning model to the drawing. This study proposed an object detection algorithm specialized for drawings by combining the non-maximum suppression (NMS) algorithm with the sliding window algorithm. YOLOv7 was selected as an object detection model, which showed the best accuracy by comparing various deep learning models. First, we made a detection window that slides on the drawing. Then, the NMS algorithm was applied to remove duplicate objects from the overall detection results. Training a deep learning model requires a large amount of training data, but it takes a lot of time to label drawings manually. Therefore, we proposed a data generation model for training data. Objects and background images were extracted from several drawings, and training data were generated by randomly mixing them as material and base. The optimal parameters for training data were selected by comparing the accuracy of the drawings. All models used in this study were trained only with the generated virtual training data. Knowing how many objects are placed in each division of the ship is important in the inspecting process. Therefore, we developed an algorithm that automatically recognizes the division of the ship and organizes the types and numbers of equipment placed in each division. Furthermore, we developed an algorithm that can obtain the connection relationship between objects and detailed specification of objects by recognizing lines and texts connected to each object. The method proposed in this study was applied to several actual plans. We confirmed the effectiveness of the proposed method by obtaining high average accuracy. By applying the proposed method, the review procedure, which took several days, can be reduced dramatically to a few minutes per drawing. |
|---|---|
| Publication Date | 2023-06-22 |
