A Multiclass Imbalanced Dataset Classification of Symbols from Piping and Instrumentation Diagrams

Engineering diagrams provide rich source of information and are widely used across different industries. Recent years have seen growing research interest in developing solutions for processing and analysing these diagrams using wide range of image-processing and computer vision techniques. In this paper, we first, present a new multiclass imbalanced dataset of symbols extracted from Piping and Instrumentation Diagrams (P&IDs). The dataset contains 7,728 instances representing 48 different types of engineering symbols and it is considered the first of its kind in the research community. Second, we present a new method for handling multiclass imbalance classification based on class decomposition by means of unsupervised machine learning methods. Experiments using Convolutional Neural Networks showed that using class decomposition significantly improves the classification performance that can be achieved, without causing information loss, as it is the case with other class imbalance data sampling approaches.