To address the issues of complex structure, low detection accuracy, and poor real-time performance in existing steel defect detection models, the authors proposed a lightweight and efficient steel defect detection algorithm based on YOLOv12n, dubbed AMD-YOLOv12n. The method first introduced Adaptive Rectangular Convolution (ARConv), which dynamically learned the height and width of convolution kernels to generate rectangular kernels, thereby effectively capturing features of defects at different scales in images. Subsequently, a Multi-Receptive Field Feature Interaction (MRFFI) module was introduced to achieve multi-scale feature fusion through heterogeneous processing of channel dimensions. Finally, the feature interaction module C3K2_DIA (Deformable Interactive Attention, DIA) was reconstructed, reducing the number of parameters and computational load while efficiently extracting features. Experiments were conducted on the NEU-DET steel defect dataset. Compared with the YOLOv12n algorithm, the AMD-YOLOv12n algorithm increased mAP and recall rate by 3.2% and 3.6%, respectively, achieved a detection speed of 260.4 frames per second, and reduced the number of parameters and computational load by 0.59M and 1.4G, respectively. This meets the dual requirements of high precision and real-time performance in industrial quality inspection scenarios.