Imaging to Perfection

No matter how hard we try to avoid the occurrence of faults during production, they are a statistical reality. Although defect rates are usually low, the stakes are high. In 2016, Samsung released the Samsung Galaxy Note 7, a smart-phone that soon proved prone to overheating. Over 100 cases of flames and explosions were reported, which led to the recall of all 2.5 million devices. Samsung, allegedly driven by the fierce competition among tech companies, performed bad quality control that ultimately led to losses amounting to over $3 Billion, not including the investment required to rebuild its image in the future. Even if faults are kept at low levels, quality seems to have an unbalanced effect on customer satisfaction, amplifying the damage of negative experience: a satisfied customer tells 6 people about a positive experience, while an unsatisfied spills to 14.

Fault detection has evolved alongside the industrial revolutions and the means of applying/exercising quality control have improved in terms of performance, by studying industrial processes and understanding the generation process of faults. But, up until now, fault detection required a human-in-the-loop, to constantly monitor a production process. Since the advent of vision inspection and image processing algorithms, cameras are now able to search for faults within the products, but the list of faults was predefined and, thus, restrictive.

Today, image recognition algorithms, which employ both image processing and deep learning techniques, can “learn” the concept of a flawless product and recognize products that deviate from it in multiple ways. Apart from the apparent advantage of reducing the cost of production by freeing up human resources from the tedious task of monitoring, this level of automation can help the industry achieve something beyond the reach of traditional, human-operated factories: omnipresent and omniscient quality control.

We should not regard this as yet another improvement in the accuracy of quality control. Instead, a complete automation of industrial processes, through a combination of advanced physical systems and image recognition that will help eliminate the parameter that has always introduced uncertainty into the evaluation of all our operations: the human factor.  To embrace this possibility, one needs to look into the power of our current technological progress.

What is image recognition?

From an AI perspective, image recognition belongs to the area of pattern recognition and is an example of supervised learning. While image processing refers to a variety of transformations that can take place to convert the image to a format useful for an application, image recognition is the technology concerned with understanding the content of the image. It primarily employs deep learning algorithms, but processing techniques from traditional image analysis are often required. This is because the input to such a system, which consists in pictures, that from a computer perspective are two-dimensional tables containing raw pixel values, often needs pre-processing: Is there noise in the picture? Is color relevant to the task? What is the texture of the objects?

Specifically in fault detection, the learning task is to find the properties of a flawless product by observing many pictures of it, provided by human specialists. This knowledge is referred to as the model, which has the form of an artificial neural network. The described process, termed training, is the computationally intensive part of deep learning and is usually performed offline, requiring time, resources and repositories of millions of images. Subsequently, the model can be easily employed to monitor newly produced products and classify them as flawed or flawless. For image recognition tasks, a specialized type of artificial neural network is used, called a convolutional neural network.

How do convolutional neural networks “see”?

Convolutional neural networks aim at extracting the useful features in an image. They consist of consecutive layers of neurons, with each layer gathering information at a specific level of detail. Simplistically, we can imagine the first layer as seeing pixels, the second seeing lines, the third abstract shapes and the fourth one objects. This convolutional network is followed by another network, termed as a classifier, whose objective is to classify the picture as flawed or flawless based on the features extracted by the previous network.

Although quite similar to classical feedforward networks, this family has the property of translation invariance. This suggests that, if the network has learned the concept of a “chopstick”, it can recognize it regardless of its specific position or angle of it in the picture. Although this ability may sound quite trivial for a human, it was not present in neural networks until recently, making the application of AI in vision-related tasks impossible.

Despite its immense power, deep learning comes with limitations. While deep neural networks can extract features and learn important concepts without requiring human intervention, there is one thing that they today need more than ever: data. This has given a number of headaches to AI experts:

• Data collection is tough, with toughness varying across industries. In some cases, data are provided by the users of an online service, such as Google maps, in which case they are abundant. But, in industrial applications, experts are traditionally required to provide labeled data, a process that consumes significant amounts of time and money. When it comes to applications associated with human safety, such as drug testing, data collection becomes an even greater concern. Naturally, this problem has not passed unnoticed from the AI community, with researchers looking for ways to get a maximum exploitation of existing data. This practise is termed as data augmentation and can take simple forms, such as applying geometrical transformations on an image, to more sophisticated ones, based on conditional GANs, which are capable of transforming images across different domains and styles.
• Training takes a long time. The larger the repository of labeled data, the longer the training times. The recent hardware acceleration that we experienced facilitated the bloom of deep learning. However, experts have shared concerns that we have over-relied our success to hardware and neglected the learning algorithms themselves. This would not necessarily be bad, if it were not for the alarming deduction of industry experts that Moore’s law, which foresaw an exponential increase in computer power, is no longer on our side.

Although the combination of cyber-physical systems and AI makes for a formidable combination toward automation and advancing quality control, we still have some way to go. A critical observer could comment that we have replaced our need for human intuition with, an easier to satisfy, yet still pressing, need for data. With the next AI discovery being just around the corner, the critical observer cannot be but receptive and vigilant.