The combination of automation equipment and intelligence has become the trend of the automation industry. Among them, automated optical inspection (AOI) applications are applied in the field of automation due to the advantages of not being limited by working hours, rapid detection, and low labor costs. Because the content of the sample database is directly proportional to the detection accuracy, so in advance the established sample library becomes an important part, and the insufficient number of training samples are the main cause of poor detection accuracy. This research is integrated with industrial practice. In the automatic feeding system, the goal of high-accuracy product detection is achieved through the method of generative adversarial network (GAN). In this study, two kinds of generative adversarial networks are used to generate unknown flaw samples and import them into the database to improve the detection accuracy. In order to simulate the insufficient number of training samples, this paper uses a small number of samples for generation experiments. In addition, this study set up an automatic feeding system, using Ethernet communication architecture to integrate six-axis robotic arm, programmable logic controller (PLC), industrial computer and other equipment, with intelligent functions such as defect detection and so on. Experimental results show that the addition of GAN to expand the data set can improve the accuracy of sample detection. In addition, the graphical user interface (GUI) is also designed in the system to reduce the learning cost and operation difficulty of personnel and improve the utilization rate of equipment, thus forming a highly efficient intelligent mechanical system.