Cold heading steel top forging cracks are challenging to capture due to their complex shapes and varying structural dimensions. Additionally, differences in understanding standards and reference images can lead to inconsistent judgment results in manual inspection. This paper introduces a detection and classification algorithm for cold-heading steel top forging cracks based on an enhanced YOLOv8n model, incorporating multiple attention mechanisms to address these challenges. The proposed algorithm first integrates receptive field attention convolution (RFAConv) and CBAM (concentration-based attention module) into the backbone network to improve the convolutional modules. Moreover, DAttention is added to the backbone network for dynamically allocating attention weights and speeding up model training. In the neck portion of the network, C2f and Conv modules are replaced with improved C2f_RFCBAM and RFCBAMConv modules to optimize the model and further increase detection precision. Finally, a new loss function, SioU, is introduced to accelerate convergence and improve regression accuracy. Experimental results show that the improved YOLOv8n model applied to a self-annotated cold heading steel dataset, achieves increases of 7.3% in mAP@0.5, 2.3% in mAP@0.5:0.95, and 10.3% in recall rate.