Classification and identification of steel scrap with deep learning

Following the industrial revolution, the rapidly increasing demand for steel has heightened the risk of depleting natural resources and intensified environmental impacts. In this context, the recycling of steel scrap holds significant importance for sustainable production. Particularly, the use of steel scrap in Electric Arc Furnaces (EAF) stands out as an environmentally friendly production method due to its high energy efficiency and low carbon footprint. However, the accurate and rapid classification of scrap to be used in EAFs is a critical requirement in terms of both process efficiency and the quality of the final product. In this thesis study, a system was developed to automatically classify scrap-laden trucks during their passage over the weighbridge. The developed system utilizes deep learning-based image processing techniques to automatically determine the scrap class and communicates the classification result to truck drivers via an LED display positioned on-site. In this way, human errors inherent in manual classification processes are eliminated, and operational efficiency is achieved in scrap logistics. The dataset used in this study was collected under real field conditions. The custom dataset consists of 11,787 labeled images representing 8 different scrap classes, with 1,407 of these images reserved for testing purposes. The fact that the data was collected directly from an industrial environment enhances the applicability of the model in real-world production scenarios. Among the models tested, ConvNeXt Large achieved the highest classification accuracy. This model predicts the type of scrap with high precision using images captured by the weighbridge camera and directs the truck to the appropriate dumping area among a total of 36 unloading points located in two separate halls (A and B). With this system, the scrap acceptance process has been fully automated, ensuring continuous and efficient operations. The findings demonstrate that deep learning-based image processing techniques offer high accuracy and reliability in scrap classification. Consequently, energy efficiency in steel production processes is increased, and production quality becomes more consistent. Future studies will aim to expand the dataset with more images and additional classes, fully integrate the system into real-time production environments, and develop hybrid models supported by multi-sensor technologies.