Abstract: The grade determination of scrap steel is a critical step for achieving efficient recycling of steel resources. To address the limitations of existing methods, such as insufficient detection accuracy and low efficiency, this paper proposes an intelligent scrap steel determination algorithm based on improved DeepLabv3+ convolutional neural network. The algorithm incorporates a coordinate attention block hybrid attention mechanism after the atrous spatial pyramid pooling (ASPP) module and replaces some dilated convolutions in the ASPP with deep strip-shaped dilated convolutions. A comprehensive dataset of scrap steel pile images under various real-world conditions is constructed, including different material types, perspectives, and time periods, to train the intelligent determination model. The improved algorithm significantly enhances detection accuracy. In control groups using ResNet as the backbone network, the m_IoU increased by approximately 2.54%, while with Xception as the backbone, the m_IoU improved by about 4.42%, effectively boosting the semantic segmentation precision for scrap steel. A conversion model based on thickness and distance factors was established to transform the pixel proportion occupied by various types of scrap steel in images into actual weight proportions. A fully connected network was employed to align the algorithm's output with manual worker annotations. Extensive experiments demonstrate that the classification accuracy of the model in this paper reaches 93.75%, significantly outperforming existing methods and meeting practical production requirements.