Single Line Electrical Drawings (SLED): A Multiclass Dataset Benchmarked by Deep Neural Networks

• Main Authors: Bhanbhro, H., Hooi, Y.K., Zakaria, M.N.B., Hassan, Z., Pitafi, S.
• Format: Conference or Workshop Item
• Published: 2023
• Online Access: http://scholars.utp.edu.my/id/eprint/37989/; https://www.scopus.com/inward/record.uri?eid=2-s2.0-85178020733&doi=10.1109%2fICSET59111.2023.10295140&partnerID=40&md5=c4fb1a4b9990d4c1b1e1a53a66483043

Single-line drawings have diverse applications across various industries, including electrical substations, buildings, power distribution, maintenance, and more. Analyzing and interpreting these diagrams using deep neural networks requires the creation of large datasets, which poses a challenging task. The complexity of these diagrams, combined with the limited availability of publicly accessible datasets, makes dataset creation even more difficult. The quality of the dataset significantly impacts the classification accuracy, as weaker datasets can negatively affect the overall performance of the model. To address these challenges, we introduce Single Line Electrical Diagrams (SLED), a multiclass dataset consisting of 3,078 instances of engineering symbols. These symbols are extracted from intricate technical drawings known as Single Line Diagrams (SLDs). The SLED dataset is meticulously curated by annotating and pre-processing appropriate images to represent various symbol classes present in the diagrams. In this article, we benchmark the SLED dataset using a variant of the You Only Look Once (YOLO) algorithm, specifically YOLO v5, for symbol classification. The results of image classification on this newly generated dataset are promising. However, further improvement can be achieved by addressing class distribution imbalances within the dataset. By making this dataset available to the academic community, we aim to enhance the understanding of the field and shed light on an important yet neglected problem within the industry. We provide a comprehensive analysis of the dataset's characteristics and demonstrate the performance of deep learning models on recently created datasets. Our conclusions offer insights into the potential directions for future research in this domain. © 2023 IEEE.