Dataset Annotation Converter for Training Neural Networks to Detect Defects on Various Surfaces

This paper discusses the development of a converter to transform RLE (Run Length Encoded) data into a format suitable for training the YOLO neural network, which is used for real-time object detection. The converter is applied to datasets of road surface damages and defects on metal sheets. The results demonstrate the converter's effectiveness in preparing annotations that enable YOLO to accurately detect defects on monotone surfaces. The authors applied their converter to the Severstal dataset, illustrating how to merge and organize data for training and validation. An example of a trained neural network demonstrates its ability to accurately detect small defects on metal sheets after 30 epochs of training, showing confidence levels for each identified object.