Conditional Diffusion to Enhance Performance of Object Detection in Unbalanced Data Engineering Drawings

To train deep learning models, it is necessary to have a large amount of data available, preferably with balanced classes. Models often struggle to achieve acceptable performance because unbalanced classes drastically reduce the model's generalization capacity. In the case of applying deep learning for digitizing engineering drawings (EDs) of the rail networks, rare symbols are common as some drawings are used to describe specific scenarios, resulting in unbalanced datasets. In this paper, a synthetic EDs generator is proposed to augment the database of relay-based railway interlocking systems (RRIS). The synthetic EDs are created based on RRIS schematic rules and probabilities considering the performance of the model using original data, new samples are created based on the results that need to be improved. Considering conditional diffusion, the proposed method achieves a mean average precision 12.54\% higher than the results of the model using only original data. Other compared models had limited variability in the generated samples and struggled to deal with unbalanced classes. The proposed approach overcame these problems and proved to be a promising technique for generating synthetic drawings.