Abstract: As a recyclable green resource, scrap steel is an important alternative raw material for iron ore and plays a key role in promoting the sustainable development of the steel industry. The grade of scrap is related to the production cost and quality, and its accurate classification and rating are very important for scrap recycling. The existing research on scrap classification generally has problems such as poor detection effect of small target scrap and background interference, which affect the accuracy of scrap classification. Therefore, an improved scrap recognition algorithm YOLO-SNBP for dense small targets based on YOLOv5 is proposed. Firstly, a small target detection layer is added to improve the recognition effect of the model on dense small target scrap in the acceptance scene. Secondly, the BiFormer attention mechanism is introduced to enhance the model's ability to extract small target features in complex backgrounds. Finally, the Soft-NMS algorithm is used to replace the traditional NMS to reduce the problem of missed detection of scrap due to overlap. The YOLO-SNBP model is trained and verified on the self-built scrap data set, and different detection algorithms are compared and analyzed. The experimental results show that compared with the basic model, the P(Precision), R(Recall) and P_mA (mean Average Precision) values of the YOLO-SNBP model are increased by 2.8%, 7.2% and 7.4%, respectively. Compared with the previous algorithms, the P_mA values are increased by 29.0%, 18.2% and 18.8%, respectively. It shows significant advantages in the accurate identification of scrap steel, and provides effective support for the efficient rating during acceptance.