To solve the problems of low detection accuracy of steel defects, large number of model parameters, and difficulty in detecting multiple-sized targets, a steel defect detection algorithm based on an improved YOLOv7-tiny network was proposed. The method first introduced GSConv(group shuffle convolution) module to reduce the redundant computations caused by standard convolutional blocks; further the multi-scale feature fusion of the model was further enhanced by embedding BiFPN(bidirectional feature pyramid network). The SPPEC(SPPELAN-CBAM) module was designed to increase the detection speed and enhance the information extraction capability; the K-means++IoU(intersection over union) algorithm was employed to redefine new anchor boxes adapted to steel defects, thereby reducing the missed detection rate. Experiments on NEU-DET dataset show that compared with the original network, the improved detection algorithm has increased the mAP(mean average precision) value by 6.8%, reduced the number of parameters to 4.71 M, and increased the detection speed to 123 f/s(frame per second), achieving more real-time and accurate steel defect detection requirements.