Deep Learning Industrial Machine Vision Processing

　　In 2012, the University of Toronto made a breakthrough in the ImageNet test performance using a convolutional neural network model trained by deep learning for the first time, which triggered a series of optimizations based on convolutional neural networks and continued to significantly improve the ImageNet test performance. In 2015, the error rate of the convolutional neural network model trained by deep learning in the ImageNet test performance had dropped to 2.3%, surpassing the recognition accuracy of humans, thus promoting a wave of large-scale industrial applications of deep learning in the field of image recognition.

　　(Deep learning technology surpassed humans in 2015)

　　Starting from 2012, major Internet giants began to implement deep learning technology. In 2013, Google used deep learning to perform OCR recognition of house numbers in Street View Maps; in 2014, Facebook applied its face recognition technology DeepFace based on the convolutional neural network model to its photo application on a large scale, with a recognition accuracy rate of 97.3%. Currently, deep learning is widely used in the fields of face recognition, self-driving cars, etc., and has achieved very remarkable results. As a technology that was first widely used in the consumer Internet field, can it be effectively integrated into the industrial vision field?

　　What tasks does industrial machine vision need to handle?

　　As an industrial automation technology based on 2D or 3D camera sensors, machine vision has extensive and mature applications in the field of industrial vision. 3C, semiconductor, automotive and other industries use machine vision technology extensively for abnormality recognition, label recognition, material positioning and other tasks.

　　Industrial automation cannot do without sensing technology and motion control technology, just like humans cannot do without eyes and hands. 80% of the information obtained by the human body comes from vision, so it is conceivable that visual sensing technology must be one of the most important technologies in the field of industrial automation.

　　However, traditional machine vision has obvious limitations.

　　What are the limitations of traditional machine vision?

　　The working principle of the traditional machine vision image processing system can be simply understood as follows:

　　1. Find artificially defined target features such as edges and corners in the image;

　　2. Perform logical judgment based on the existence of target features in the image and the numerical value of the distance between multiple target features to complete the visual task.

　　To use this technology, vision engineers need to define target features and thresholds for numerical judgment based on the specific requirements of the vision task. Once designed, they form a program to be executed by the machine.

　　The limitation of traditional machine vision logic is that it cannot be applied to tasks with strong randomness and complex features. Typical tasks include:

　　(Detection of random complex appearance defects)

　　Since only a limited number of features can be arranged and combined, visual engineers cannot express comprehensive and complex judgment targets such as "dense dot-like unevenness" through "edges" and "corners". Or the expression ability is very poor, resulting in poor recognition accuracy. Therefore, traditional machine vision cannot solve the above problems.

　　Such complex feature problems are exactly what deep learning technology is best at solving.

　　How does deep learning solve complex feature problems?

　　Compared with traditional machine vision, where vision engineers design algorithm models, the biggest difference of deep learning technology is that the program can independently discover what features are needed and what logical relationships are needed to complete image analysis tasks, thereby enabling the program to design algorithm models.

　　To use Lego blocks as an analogy, in traditional machine vision, humans pick out dozens of Lego elements from 100, assemble them to perform the logical actions designed by humans, and complete the relevant tasks; while in deep learning, humans tell the machine the tasks to be completed, and the machine picks out tens of thousands of Lego elements from 100 million, assembles them, and selects the logical actions to be performed to complete the task. Its expressive ability is far higher than that of human experts.

　　(Deep learning technology has expressive capabilities far beyond those of human experts)

　　Since deep learning can choose from more possible features and determine the logical relationship between features on its own, deep learning has the ability to express 'dense dot-like bumps and bumps' by selecting a set of features from a large number of pixels.

　　(Features and logical combinations are selected by the software independently)

　　In practical applications, models trained using deep learning can accurately identify random defects in images and effectively identify specified defects, truly realizing the detection of random defects with strong randomness and complex features.

　　(Deep learning can identify and label random defects in images)

　　It is precisely because of its ability to handle such image recognition problems with high randomness and complex features that deep learning has the potential to break through the limitations of traditional machine vision technology.

　　Can deep learning achieve industrial precision?

　　We usually think that industrial applications have higher requirements for technical accuracy and stability than civilian technologies. So, can the deep learning technology that is popular in the consumer field meet industrial indicators? Let's take appearance defect detection as an example to see what specific indicators need to be considered in industrial detection.

　　(Accuracy assessment matrix for detection tasks)

　　Missed detection rate: Missed detection will directly cause defective products to flow to end customers. Therefore, the missed detection rate requirement is usually less than 100 PPM.

　　False positive rate: False positives will directly affect the yield of industrial enterprises and cause waste of materials. Enterprises usually require the false positive rate to be between 1% and 5%. On the premise that the missed positive rate meets the standard, only by significantly reducing the false positive rate can the goal of reducing manpower be achieved.

　　Beat: There are great differences in different industries. For example, the beat requirement in the electronics industry is within 5 seconds, while the beat requirement in the mechanical processing industry is within tens of seconds.

　　On the one hand, the current industry-wide technical level of deep learning has been able to achieve a judgment accuracy rate of more than 95%. By balancing the missed judgment rate and the false positive rate and more strictly controlling the missed judgment, the missed judgment rate can be reduced to below 100 PPM and the false positive rate to below 5%.

　　On the other hand, in terms of the beat requirement, since the current GPU graphics card can achieve an image processing speed of 80 frames per second, 400 pictures can be judged within 5 seconds. Generally, the products in the 3C industry are small, and only 10 or less pictures are needed to complete the product coverage. For example, large machined products only need less than 100 pictures to fully cover the product surface. The speed of image processing can meet the beat requirement.

　　So overall, we believe that deep learning technology has matured to the point where it can accomplish complex industrial vision tasks.

　　In fact, deep learning has been productized

　　Yes. UnitX is based on such a technical judgment and integrates deep learning technology into the field of traditional machine vision to solve the problem of complex surface defect detection. At present, UnitX has successfully realized productization in the fields of complex machined product appearance defect detection and high-reflective plastic product appearance defect detection. The detection effect is much better than that of traditional visual inspectors. It has completed more than 300,000 materials without missing any inspections, and the inspection cycle has increased by 40%. It has realized the automation of appearance defect detection, proving that deep learning technology can meet industrial inspection needs.