With the accelerated progression of industrialization in China, the annual volume of decommissioned steel has surged significantly, resulting in the accumulation of massive steel scrap resources that form “urban mines”. These waste materials not only occupy substantial land resources but also pose potential environmental threats. In this context, short-process steelmaking, which utilizes electric arc furnaces to remelt steel scrap for new steel production, transforms scrap into a green and eco-friendly “industrial bloodstream” . Within this process chain, accurate and efficient evaluation and acceptance of steel scrap constitute a critical link for ensuring production efficiency and cost control. However, current manual inspection methods exhibit inefficiency and subjectivity, thereby constraining production scalability. Furthermore, valuation discrepancies among different inspectors can reach up to 300 yuan per ton, increasing the risk of transaction disputes and integrity violations. In present study, we report a hierarchical scene-based recognition framework and propose a multi-dimensional temporal deduplication algorithm for steel scrap, aiming to effectively address background interference, achieve precise deduplication, and maintain inter-frame consistency. The proposed method demonstrates significant advantages in terms of deduplication granularity, accuracy, and robustness, providing critical technical support for high-precision automated grading of steel scrap. We anticipate that this approach can be further extended to intelligent lifecycle management of steel scrap and contribute to establishing a traceable dispute resolution system for scrap recycling. This advancement is expected to facilitate the digital-intelligent transformation of the scrap recycling industry.

This paper proposes an intelligent steel scrap grading framework based on temporal deduplication technology, comprising four progressive operational phases. Initially, the system automatically extracts keyframe images from unloading videos by analyzing spatial relationships between grab buckets and cargo compartments. Subsequently, panoptic segmentation is applied to each frame to obtain category-agnostic steel scrap instances, followed by a bidirectional association algorithm for cross-frame object ID assignment. The backward association matches current-frame detections with prior-frame instances to inherit or allocate new IDs, while the forward association retroactively corrects prior-frame associations using current-frame IDs. The SAM2 tracker enhances temporal consistency through its visual similarity computation module, where memory encoders and attention mechanisms effectively address target morphological variations. Next, a change detection algorithm localizes dynamic regions within the compartment to filter background-interference IDs, thereby focusing statistical analysis on valid areas. Finally, a hybrid material-type recognition algorithm classifies retained instances. A composite segmenter, integrating EfficientViT-SAM, first generates target proposals via FastSAM and then refines mask segmentation. Material type proportions are statistically derived through ID-indexed recognition results. This temporal deduplication mechanism synergizes bidirectional association with composite segmentation, reducing inter-frame redundancy while ensuring spatiotemporal continuity in steel scrap counting, thereby providing a robust solution for industrial material grading applications.

This paper presents comprehensive experimental evaluations to validate the proposed methodology. First, the designed Area-of-Interest (AOI) attention module was tested in smoke-occluded scenarios. Without AOI integration, detection results suffered significant smoke interference, achieving only 48% Intersection over Union (IoU) in changing regions. In contrast, AOI-enhanced implementation attained 96% IoU (Fig. 9), demonstrating exceptional robustness. Second, critical threshold parameters in the bidirectional association algorithm were optimized using Identity Switches as evaluation metrics, yielding optimal combinations of threshold parameters σb=0.1, σf=0.9 and σa=0.1. Subsequent visual analysis of algorithm stages (Fig. 10) reveals that the complete bidirectional association effectively integrates complementary advantages of backward and forward associations, ensuring unique ID assignment and precise temporal visualization for identical targets. Furthermore, real-time performance evaluation of the multi-dimensional temporal deduplication method quantified computational efficiency improvements. Optimized backward and forward associations reduced processed object counts from 290 to 50 and 240 to 70 respectively, with total execution time decreasing from 17.7 s to 7.4 s,notably below the 21.2 s average grab cycle duration. Finally, comparative analysis using relative error metrics demonstrates the superiority of our method, achieving merely 5% error versus conventional approaches [direct counting (160.7%), NMS (72.0%), LASER (41.1%), standalone backward association (67.3%), and standalone forward association (11.2%)]. These results conclusively establish the proposed method’s state-of-the-art performance in long-sequence scrap temporal deduplication tasks.

Current intelligent steel scrap research predominantly focuses on “quality inspection” tasks, while the critical temporal deduplication problem in “grading” scenarios remains understudied. This paper presents the first systematic exploration of scrap temporal deduplication through a hierarchical scene recognition framework. Addressing multiple challenges including background interference, precise deduplication, and temporal consistency, we propose a multi-dimensional temporal deduplication algorithm for scrap materials. Experimental results demonstrate that our method exhibits significant advantages over existing technologies in deduplication granularity, robustness, and accuracy. Despite these advancements, the algorithm’s performance remains partially dependent on the underlying object tracker’s effectiveness. Particularly when handling multiple targets with high visual similarity, the tracker may generate unstable temporal associations (e.g., temporal ID flickering observed in the red truck side-rail case shown in Fig. 11), thereby compromising the robustness of final temporal deduplication. Future work will prioritize resolving precise deduplication challenges involving multiple visually similar objects to enhance system reliability.