Research on Scrap Steel Identification and Detection based on SAM and YOLOv9

Abstract

In the context of the "dual carbon" initiative, the steel industry must expedite its efforts to enhance the automation and intelligence of scrap steel recycling, as this is a pivotal step in achieving a low-carbon transformation and ensuring resource circularity. The prevailing methods, which depend heavily on manual labor, are characterized by inefficiencies and suboptimal accuracy. In response to these challenges, this study proposes an innovative approach that integrates the Segment Anything Model (SAM) with the YOLOv9 object detection algorithm for scrap steel classification. Initially, SAM is employed to meticulously delineate scrap steel images, effectively isolating the target regions from their complex backgrounds. Subsequent to this segmentation, the YOLOv9 algorithm is implemented for real-time detection and classification of the segmented regions. Empirical evaluations demonstrate that, in comparison with conventional YOLO-based methodologies, the proposed approach enhances mAP@0.5 by approximately 7% and elevates the recall rate by around 2%. Furthermore, the method demonstrates enhanced robustness and precision when handling scrap steel images with varied morphologies and intricate backgrounds. These results underscore the promise of a segmentation-first strategy in advancing the automated classification of scrap steel, thereby paving the way for more efficient industrial recycling processes.