Enhancing Classification in Imbalanced Symbol Engineering Drawings using Affine 2D Geometric Transformation

Abstract

In the design of smart cities, the digitisation of engineering symbols ensures that accurate intelligent systems can be deployed. Errors, resulting from the inability of humans to accurately read and analyse manual engineering symbols often lead to catastrophic consequences. However, the digitised engineering symbols come hampered with the class imbalance problem. Some recent publications have addressed the detection and classification of image symbols without considering the class imbalance issue. This has resulted in misleading predictions of only the majority class instances. This paper proposed affine 2D geometrical transformation technique of angle rotation, vertical and horizontal flipping, shearing, and cropping to augment the instances of minority classes in symbols in engineering images (SiED). The instances of minority classes were first transformed using the 2D geometrical transformation then the symbols of the minority classes transformed to generate more artificial instances to balance the class distribution in the dataset. The proposed work is evaluated with the "Symbols in Engineering Drawing" (SiED) Dataset. Convolution Neural Network (CNN) algorithm was used to evaluate the proposed model based on accuracy, balanced accuracy, precision, recall, kappa and training time. The affine 2D geometrical transformation image augmentation method recorded a high performance in all the performance metrics.