Artificial Intelligence: Understanding How Object Detection Creates Business Value for Manufacturers

See the new AI tools that are disrupting manufacturing operations.

The significant impact that Industry 4.0, IoT, and Smart Factories are expected to have on manufacturing this decade is well documented in reports by numerous industry experts and leading advisory and consulting firms. A common observation across these reports is that maximum impact will primarily be derived through efficiency and productivity gains enabled by applications of artificial intelligence and through unique insights extracted from big data. The high-level use cases where artificial intelligence and big data will facilitate many of these improvements — intelligent maintenance, quality assurance, and inventory demand and sales planning— are well documented and clearly outlined, but for readers without specific knowledge of the workings of AI or unfamiliar with how AI functionally operates at specific tasks, there may remain a somewhat incomplete understanding of the value delivery of AI. The intent of this article, the first of a 3-part series on AI and manufacturing, is to illustrate a number of those task scenarios through an introduction of object detection, a sub-field of computer vision and artificial intelligence, and outline how the application of object detection uniquely creates value within manufacturing industries.

Let’s first come to a quick understanding of what object detection is. Object detection can be defined as finding, locating, and recognizing objects within an image. With simple object recognition, an image is analyzed by AI and given a classification label. The label could be a descriptor that identifies the object of interest within the image. For example, “a car” or “an airplane”, or it could relate to a quality determination — “pass” or “fail”. With more complex object detection, the object’s location within the image is also identified, and a unique bounding box with a classification label is created to identify that object. A busy image containing multiple identified objects will see each individually identified object receive a bounding box and classification label.

With simple object recognition, AI performs analysis of a captured image and outputs a classification label. With more complex object detection, objects within an image are identified, located, and issued individual bounding boxes with classification labels.

To build on object detection functionalities even further, masks can be applied to the image so that located objects are identified and grouped at a pixel level, with each object grouping visually represented by a different color. This pixel-level representation of objects not only improves accuracy in sizing, measurements, and area calculations of objects, but it also improves the very critical ability to perform the initial identification and localization of objects within an image.

With instance segmentation, there is a pixel-level identification and classification of identified objects within an image. Each identified object is visually represented by a unique color.

Learning what object detection is and how it can be applied is certainly a valuable insight. Realizing what it actually enables though, will provide far greater value to organizations considering its implementation. Specifically, object detection has the ability to take raw unstructured image and video data, analyze it, and transform its output as organized structured data. Data in image or video format is fluently interpreted and efficiently converted to accessible operational information in the form of product quality evaluations, inventory numbers, production bottlenecks, employee safety, traffic patterns within a facility location, and more. Object detection is in fact going to be a facilitator of enhanced data insights for image output and streaming videos, and it will be immensely valuable across manufacturing industries with the following target applications.

Quality Assurance — Surface Defect Inspection

Musashi, a global supplier of suspension and power train components, estimates that it spends 20% of its time and costs on ensuring that the quality of its finished goods meet the requirements and expectations of its end customers. For organizations facing these same challenges, an ability to substantially decrease those time and financial costs will undoubtedly have a positive impact on the company’s bottom line and has the potential to create a source of competitive advantage as the employees required to complete the very laborious and repetitive task of final goods inspection can be re-positioned within the organization to perform greater value added activities.

Although there has been industry adoption of automated optical inspection processes for some simple inspection tasks, manufacturers of complex parts with challenging inspection requirements still tend to rely on manual inspectors as their final quality control measure. Given that there are known performance limitations with human inspectors, and inspection skills can take significant time to develop, there are clear opportunity factors where despite best efforts, defective goods may still reach an end customer. A quality failure like that can lead to costly customer issues, product recalls, warranty claims and damaged brand reputation.

In fact, the American Society for Quality (ASQ) estimates the cost of quality failures in many organizations is as high as 15% to 20% of sales revenue. Within competitive supply chains, companies battle fiercely for supplier contracts to large OEMs, and a good majority of those engaged in the battle view the final quality of their goods as an absolute competitive priority. As an upstream partner, they are under immense pressure to produce and ship goods of superior quality in order to win and retain those critical supplier contracts. Companies considering deployment of AI to perform visual inspection, could realistically see a 90% improvement in defect detection and a 50% improvement in manufacturing productivity.

Video of object detection with deep learning. Watch defect detection with classification and assembly verification being performed on a machined metal part.

As seen in the video above, the improvements in defect detection and productivity are mainly due to the observable fact that object detection is able to inspect complex parts for microscopic defects as they come off a production line with better accuracy and consistency over human inspectors. And, it achieves that performance improvement while automating data capture processes and increasing the amount and quality of the data being collected.

The data insights enabled by object detection will directly contribute to productivity improvements for manufacturers in a number of ways. When an organization can reliably access the real-time identification and classification of defective goods, root cause analysis and prevention of those defective goods becomes highly manageable. Through improved responsiveness to quality issues and adaptive corrective measures, costly waste in the form of scrap and rework is minimized or completely eliminated. Additionally, by automating data capture processes and by enabling better data insights, quality departments will receive crucial support and enhanced tools for important activities like quality process audits and supplier performance monitoring.

Assembly Verification and Completeness Checks

In addition to inspecting finished goods for surface or cosmetic defects, another critical activity for quality control departments is detecting products that have been improperly assembled or those that are incomplete and missing necessary components.

Subtle feature differences between components can be accurately identified with AI through object detection. An assembled or finished product can be rapidly inspected to ensure that it contains the right sub-components before being shipped to the end customer.

A major advantage that deep learning based object detection has over traditional machine vision systems is that the AI performs exceptionally well in the fast-paced and harsh manufacturing environments where the reliability of the vision system is a crucial design requirement. Changes in lighting or illumination, the physical position of the part, or other environmental noise and conditions cause traditional vision systems to perform with lower inspection accuracy or report a problematic high number of false positives. AI, on the other hand, is able to ignore or adjust to these environmental factors, and it can perform reliable and consistent inspection in milliseconds as parts or finished goods travel down a production line.

Actual images from a vision system on a production line that utilizes object detection’s feature recognition and measurement capabilities. This vision system is being used to verify that the bushing component presented is the correct one required prior to its assembly into a control arm.

Other Applications of Object Detection in Manufacturing

Employee Safety — Deep learning based object detection can be used to actively monitor at-risk employees to ensure that the proper protective safety gear is being worn in areas where it is required. Furthermore, a smart vision system could be designed to monitor restricted access areas or give immediate warnings to employees on the shop floor when a potentially hazardous or dangerous situation arises. The National Safety Council estimates that the total cost of workplace injuries in the USA is around $170 billion per year.

Inventory Management— With object detection’s ability to accurately identify products by location, size, color, or feature, tracking and monitoring inventory items in real-time through automated AI processes is readily achievable. Production disruptions can be minimized or prevented through a smart purchasing system suggesting or executing purchase orders based on known inventory levels and anticipated consumption.

Robotics and Automation — As the costs of industrial robots fall year after year, their integration into manufacturing facilities and production lines continues at an accelerated pace. Automation is expected to boost worker output up to 30%, and lower manufacturing costs by 20–25%. Typical uses for these robots include goods handling with robots loading parts and components from pallets, bins, or containers into machinery on the production line, or extracting finished goods from the line and completing packaging tasks. Bin picking or pick and place routines are commonly utilized to accomplish these tasks and the execution of those routines is primarily dependent on vision systems that employ object detection.

Update — We’ve launched an iOS App!

You can now trial Musashi AI’s object detection capabilities with an Apple iPhone or iPad device. Download our free iOS app, MAI Cendiant^® Inspect Mobile from the Apple App Store here. The app has four trained object detection models embedded for users to trial and all AI processing is done on your Apple device (edge AI).

Engineers at Musashi AI have designed and developed a number of proprietary object detection and instance segmentation algorithms and model architectures that truly expand the capabilities of computer vision and AI visual inspection for manufacturers. These algorithms and models have been used to perform complex defect detection, feature presence confirmation, assembly verification, and process completeness checks across a wide variety of applications and finished goods.

Musashi AI believes that real value is delivered to customers by supplying custom-crafted deep learning solutions that achieve an incredibly high degree of accuracy while requiring far less image data upfront than typically needed for AI development. Improved performance at lower development costs translates into incredible value for our customers.

If you’d like to learn more about Musashi AI’s deep learning capabilities or our automated inspection solutions, please visit our website here.