The high proportion and large-scale variation of small targets with defects caused by the complex structure of metal gear end faces have led to low detection accuracy, making it difficult to meet the real-time online detection needs of enterprises. In this paper, we propose a metal gear end face defect detection method based on an adaptive multi-scale feature fusion network (YOLO-Gear) using the YOLOv5s network. Firstly, we establish a gear end face defect detection test platform and create a gear end face defect dataset. Then, we introduce the adaptive convolutional block attention module (CBAM-C3) which combines channel attention module ( CAM) and spatial attention module (SAM) to enhance the adaptive feature learning and extraction for small target defects in metal gears, effectively improving the detection accuracy of the model for small target defects. Finally, we propose the bidirectional feature pyramid network (BiFPN), which repetitively weights and fuses features from different scales, thereby improving the model’s ability to detect defects at multiple scales. Experimental results demonstrate that the YOLO-Gear model achieves an average precision of 99. 2%, an F1 score of 0. 99, and an FPS value of 33 on the gear end face defect test set. Compared to other deep learning models, the proposed YOLO-Gear model in this paper improves both detection accuracy and efficiency, meeting the real-time online detection needs of enterprises.