CoT-Integrated Lightweight Cross-Modal Model for Scrap Steel Recognition

Under the strategic goals of "Dual Carbon," the green transformation of the iron and steel industry hinges on the efficient recycling of scrap steel resources. To address the limitations of existing scrap steel recognition methods—such as low efficiency, reliance on manual experience, insufficient unimodal analysis capability, and deployment challenges—this study proposes a novel cross-modal recognition large model that integrates chain-of-thought (CoT) reasoning and lightweight knowledge distillation. Firstly, a cross-modal recognition framework based on CLIP and SAM is constructed to establish a "shape-image-composition" semantic mapping, enabling fine-grained segmentation and compositional association of scrap steel. Secondly, a model compression strategy incorporating multi-dimensional knowledge transfer and chain-of-thought guidance is designed. This strategy effectively adapts the capabilities of the large model for edge computing devices while preserving high accuracy and ensuring the interpretability of the decision-making process. Finally, an intelligent decision-making closed-loop system of "composition prediction - charge optimization - process calibration" is developed by integrating the aforementioned model. Experimental results demonstrate that the optimized lightweight model achieves an inference speed of 35 FPS on edge devices with a mean Average Precision of 92.1%. System-level simulation shows an 8.2 percentage point increase in scrap steel utilization rate and a significant enhancement in process stability. This research provides a high-precision, high-real-time, and highly reliable solution for the intelligent upgrading of short-process steelmaking.