In the manufacturing industry, testing is an indispensable link. Visual inspection can ensure that our products meet their expected functions and appearance, and bring important benefits to manufacturers and customers. Most importantly, the test results can provide quality assurance. We can communicate directly to customers through product labeling. If the product is returned from the customer, the test report can also help with troubleshooting and help the manufacturer deal with any claims.

In addition, identifying all nonconforming items during the production process can help determine whether manufacturing processes or procedures need to be adjusted. The test results can help determine the cause of the failure. For example, nozzles in surface mounters for electronic products are clogged, bottling equipment malfunctions or labeling machinery is not aligned, etc. Real-time identification of defects can stop production immediately and solve the problem immediately. The sooner the quality problem is found, the lower the cost of solving the problem.

From manual detection to AOI

Usually we need to test every product produced. Trained operators can perform manual inspections, especially when dealing with simple products or final inspections as a whole. Some products such as PCBA may require amplification equipment. The smallest functional size (such as high-density IC interconnection and 01005 size SMD chip soldered on the circuit board) poses a great challenge to the visual acuity of the inspector.

However, as the complexity of products increases, some typical components may contain a large number of such devices. When inspectors perform inspections and record the results, they must overcome visual challenges, which may make manual inspections impractical. In some cases, such as high-speed filling processes, manual inspections may not be possible at all.

As the challenges in feature size, complexity, and throughput become more severe, automated optical inspection has become the only practical method to ensure adequate inspection of each item.

AOI includes image sensing, lighting, and computing subsystems that work together to capture and analyze images. The AOI system can compare the captured image with the reference image, and then be able to identify defects on the surface of the material, soldering defects, or missing or misplaced components on the PCBA. Or, a system based on certain rules measures feature dimensions (such as the component itself or the amount of solder in each joint) to determine "good" (G) or "not good" (NG) status. If a defect is detected, the machine equipment can isolate the defective item and then continue subsequent inspections, or suspend and warn the operator.

Although AOI has surpassed manual detection in the presence of complexity and throughput, traditional image processing systems and algorithms still have some shortcomings. These shortcomings are very obvious during system and software development and equipment installation on the factory floor.

From traditional image processing to AI

The basic principle of image recognition is to digitize each captured image and apply various filters to detect patterns and features. Edge detection filters are usually used to detect objects in an image. Algorithms that can recognize humans can apply slope detection to recognize features such as arms, shoulders, and legs. The algorithm also needs to detect the direction of these detected features relative to each other as a further criterion for defining.

Whether it is people recognition in applications such as security surveillance or automobile pedestrian detection, facial recognition in social media applications, or defect detection in industrial inspections, traditional image recognition faces many challenges.

Defining rules and creating algorithms to detect and classify objects in digital images is very complicated. In industrial inspection, the development of reliable algorithms is expensive and time-consuming. When testing PCB components, as long as the quality of the solder joints is a standard for testing, we must verify the existence of each component. It is almost impossible to create rules for all situations and all exceptions.

To a certain extent, artificial intelligence can imitate human beings to apply the learned experience to image recognition, so as to be able to cope with the challenges brought by infinite changes. Among the various computing structures covered under the overall concept of AI, convolutional neural networks (CNN) are usually used for image recognition. These include artificial neurons that are connected to each other and arranged in layers. They are usually deep neural networks, containing multiple internal or hidden layers between the input and output layers. The hidden layer performs specific and strictly defined pooling and convolution calculations on the data received from the previous layer. The result is sent to the next layer, and finally to the output layer, which can indicate whether the object you are looking for has been identified.

Bring the two fields together

AI can bring huge advantages to AOI equipment suppliers and users. From the supplier's perspective, if AI can determine the probability of seeing a particular object, it can simplify algorithm development. By reducing the workload of defining each object and corresponding acceptable standards, it helps to shorten the time to market for new devices and reduce ongoing software support costs. For users, the enhanced AOI realized by AI can simplify the detection system setting, programming and fine-tuning the threshold of "good"/"bad" alarms.

The AOI on the production line can run at a speed that matches that of the production line, and is already supporting manufacturers in all walks of life to improve quality assurance and continuously improve production processes. Further improvement on the basis of AI is the future development direction of AOI. Algorithms trained for optical inspection applications can bring additional benefits of improved decision-making capabilities, reduce operator involvement, simplify programming, and provide more powerful performance, which can improve the certainty of defect detection while reducing false alarms.