Research on Dense Small Target Scrap Steel Type Recognition Algorithm Based on YOLO-SSNP

Conclusion

This work suggests an improved scrap shape recognition model, YOLO-SSNP, based on YOLOv5, to solve the issues of severe missed detections and restricted small-target extraction in complicated scrap shape identification settings. The model incorporates a small-target identification head to enhance the fusion of shallow and deep features, introduces the lightweight and efficient Slim-Neck module to enhance feature extraction and reduce model complexity, and replaces the conventional NMS algorithm with Soft-NMS to mitigate occlusion- and overlap-induced detection errors. The YOLO-SSNP model achieves 97.9% precision and 89.7% mAP, representing improvements of 3.6% and 7.0% over the benchmark model, and outperforms mainstream target identification models including the YOLO series, SSD, and Faster R-CNN. The efficiency of the suggested enhancement technique is validated by experimental findings showing that YOLO-SSNP performs well in recognizing tiny and highly overlapping scrap targets. It also offers dependable technological assistance for intelligent scrap detection in the steel sector.

Despite these advances, the model's capability to detect ultra-small targets and its robustness under extreme environmental conditions still require improvement. Future research will investigate the use of higher input resolutions and multi-scale feature fusion to preserve finer details, implement small-target-centric augmentation strategies, such as magnification augmentation, to improve the model's representation of ultra-small targets, and re-evaluate loss functions to address the class imbalance between small and large objects. Furthermore, this study will incorporate multi-source image acquisition, including complex interference environments such as high dust concentrations, nighttime illumination, and backlighting, to enhance dataset diversity. These approaches are expected to further improve the model's adaptability and detection accuracy in industrial scrap steel quality inspection scenarios.

Abstract

Scrap can lower production costs for metallurgical companies and enhance environmental sustainability. It is a recyclable resource and an essential substitute for iron ore as a raw material. The accuracy of classification is hampered by the limits of current scrap steel recognition techniques, which include limited category coverage, small-target detection capabilities, and resistance to background interference. This research suggests YOLO-SSNP, a dense, small-target scrap steel material detection algorithm created by improving the YOLOv5 framework, as a solution to these issues. To improve the model's ability to extract fine-grained features, a small-target detection head is first included. Second, to maximize feature representation while preserving model compactness, GSConv and VoVGSCSP are used to build a Slim-Neck module, which replaces conventional Conv and C3 modules in parts of the Neck layer. Finally, Soft-NMS replaces traditional NMS to enhance recognition accuracy for occluded and overlapping targets. The model is trained and evaluated on a self-constructed scrap steel image dataset and compared against several mainstream detection algorithms. Experimental results demonstrate that YOLO-SSNP offers superior accuracy and model compactness, with a precision (P) value of 97.9% and an mAP of 89.7%. It highlights the model's efficacy in precisely and efficiently identifying the different types of scrap steel material by improving the mAP value by 28.6% and 17.8%, respectively, over earlier methods.