D-Reader
Drawing recognition- Year:2022
- Technology:ШІ
During the planning of adaptation and enhancement of the existing technology to read pipeline diagrams, two stages of development have been defined along with problems to be solved. In the first stage, we decided to focus on porting existent engineering drawings analysis algorithms to detect and recognize separate elements. The following problems were planned to be solved until mid-term evaluation:
- Pipeline elements specification detection and OCR.
- Detection of the main pipeline and its direction.
- Localization of pipeline diagram elements, their coordinates, and the BOM entry.
The results could be combined to build a logic scheme for the pipeline drawing. This is the goal of the second stage alongside testing, technology improvements, and user interface development.
After the project was broken down into tasks, we discovered that the estimated time doesn’t fit a ten-month period, so a meeting with the local DIH was initiated to adjust the plan. Several approaches, which bring different levels of accuracy and development time, have been prepared for each aspect of the technology. As a result of the discussion, we conclude not to sacrifice the accuracy and performance of the technology but reduce the estimated time by the user interface (API) and back-end design simplification.
Average accuracy of the OCR classification model is 88% (89% on table data and 87% on element specs). Proportion of the correctly recognized BOM entries: 93%. Proportion of the correctly recognized pipeline diagram elements: 47%. Average execution time: 251 seconds.