February 2021

Development of Text and Image Extraction Technique in Drawing

05 February, 2021

LEE Jeong-youl(General Manager), KIM Jong-oh (Principle Surveyor), LEE Gap-heon (Surveyor)

1. BACKGROUND

The KR is researching ways to increase productivity by applying various new technologies such as big data and deep learning, which are technologies related to the 4th industrial revolution, to work. In recent years, as interest in work process improvement and work efficiency improvement has increased, interest in digitizing and utilizing drawings is increasing.

Previously, humans directly digitized the information by searching for text (quantity and specifications, etc.) and images (equipment materials, place signs, etc.) in drawings, requiring a lot of effort and a long time. In order to improve the efficiency of this, there is a need to extract the location and information of text and images in the desired drawing and digitize it for the purpose. However, due to the characteristics of ship drawings, there are situations where text and images in the drawings overlap each other or there is noise due to lines, so that people can only see them. For this reason, it is difficult to use the existing object detection algorithm because the recognition rate is low. Therefore, The KR has developed a character and image extraction model using a deep learning model that can operate efficiently even in such situations.

2. KEY FEATURES

· Selecting text and image information in drawings
· Object image editing/add/remove function
· Automatic generation of large amounts of image training data

Figure 1. Drawing properties and project information management screen

Figure 2. Text and image information recognition screen in drawing

3. EXPECTED EFFECTS

Through this study using algorithms based on deep learning, we have secured a base technology that can digitize information about text and images in drawings to suit the purpose. First, we applied drawing recognition technology to the safety plan work where you need to determine the type and number of specific facilities. Deep learning learning was conducted based on 200,000 pieces of learning data for 150 types of objects such as fire extinguishers, fire hoses, and gas detectors (as of 2020, IMO Resolution A.1116). As a result of applying it to the drawing, it showed high accuracy in finding facilities with an average of 85%.

The future research direction is to apply the developed model to various drawings to confirm the possibility of expansion and to improve the recognition model performance. Through this, we will endeavor to become a digital class leader in the 4th industrial revolution by providing high-quality inspections and smart support to those who visit The KR.

ICT Solution Team

KR Webzine Vol.156