Printed circuit board (PCB) is a critical component of electrical products, and its quality control during the manufacturing process cannot be overemphasized. This work proposes a model for early PCB defect discovery in the molded-interconnect-device manufacturing process. Based on transfer learning and data augmentation, a one-stage deep object detection network is built for defect detection, which is trained with data directly obtained from the production line. To demonstrate the effectiveness of the proposed method, the manual inspection process and its statistical data from a real PCB plant are used as a basis for comparison. In addition, a 10-fold cross-validation is performed to provide a more concise evaluation. The result shows that the proposed model possesses the ability to detect six types of subtle defects and achieves an identification accuracy of 83.75%. Moreover, the model provides a significant reduction in manufacturing cost, with 84% of the total inspection time being saved. With the advantages of accurate multiclass detection ability and low establishment cost, the proposed model is shown to have great potential for industrial implementation.