CNN-based Transfer Learning in Intelligent Recognition of Scrap Bundles

Scrap bundles offer numerous benefits, but their composition and quality can vary significantly, making accurate recognition of each bundle a key challenge. In this study, a new dataset called RSBL was created and the performance of popular CNN models was improved through transfer learning technology. This approach effectively enhances the recognition accuracy of scrap bundle models, and the most suitable model among the improved models for scrap bundles intelligent recognition application scenarios is determined through three sets of comparison experiments. The results demonstrated that MobileNet_V3_Large model improved by the transfer learning technique performed better in the scrap bundle intelligent recognition scenario, with one training epoch time of 62.54 s on the recognition of scrap bundles (RSBL)_10 dataset and average test accuracy of 99.5%; one training epoch time of 89.78 s on RSBL dataset and average test accuracy of 99.8%. Using the MobileNet_V3_Large model for scrap bundle recognition can improve the accuracy of the static and dynamic model in converter steelmaking.