In this study, we propose an improved YOLOv3 network detection method to solve the problem that gear defects are difficult to detect in industrial manufacturing. First, a gear defect image database is constructed by performing various activities, including image acquisition and expansion and defect labeling. Second, the feature extraction ability is improved using the DenseNet network structure instead of the original network structure. Finally, the small-size defect detection ability is improved by increasing the network prediction scale. When compared with the YOLOv3 network, the mean average precision and the missing-part precision of the gear increased by 3.87% and 5.7%, respectively, using the proposed method. This experiment demonstrates that the proposed method exhibits several advantages and that the gear defects can be effectively detected.