Product packaging has a wide variety of properties. Hence different shapes, colors and materials can cause different defect patterns. Data Spree from Berlin shows how Artificial Intelligence (AI) can be used to identify and eliminate complex packaging quality problems in real-time production quickly and effectively. ADLINK technology and Data Spree combine industry-ready hardware with the very latest Vision AI software. For this use case, ADLINK has a set of cutting-edge devices and technology. With the NVIDIA Jetson Neon Series based cameras, machine learning is allowed in the camera at the edge. Finding defects at an early stage in continuous production and logistics and eliminating them in a short time is often a very tedious task in the packaging sector. The reliable automation of visual inspection is crucial to ensuring consistently high quality. The challenges of visual inspection Classical image processing systems and traditional algorithms for visual inspection of packaging defects are often very inflexible and expensive. In this case, the defect detection feature has to be manually developed by experts, which requires a lot of knowledge and time. At the same time, many of the patterns of potential defects (tearing, missing pieces, scratches, dents, engineering deviations, missing contents, printing errors) can only be executed with great effort or without them at all. All this leads to higher costs and to the fact that the quality requirements for automation cannot often be met. Efficiently solve complex quality problems with Vision AI (AI) that can be used for the reliable detection of a wide range of individual error patterns and anomalies. With Deep Learning DS software from Data Spree, Vision AI software logic can be efficiently and easily implemented in the background. Continuous production data monitoring, automatic defect classification, and time-series analyzes are also implemented in an easy-to-use manner with Deep Learning DS in production processes. Figure 1: Deep Learning DS – Data Management and Statistical Analysis of Blister Encapsulation Surfaces for Automatic Visual Quality Control To implement AI-based flaw detection, package images are first taken from the production process. Now, some drawbacks that need to be discovered for AI training can be distinguished. Labeling the data is called an annotation. However, if one wants to spot general flaws and anomalies, AI, without tagging the flaw, can identify deviations from the norm by detecting the anomalies after training. In the process, the AI is frequently training to detect and localize deviations or anomalies from good condition or also particular error patterns that one would like to identify as a user. Here, AI works similarly to the human brain and learns to recognize, map, and localize defects based on image data – without having to pre-select specific encapsulation features manually with Deep Learning DS, you can do this learning process yourself quickly and easily. Data Spree also offers the complete process up to productive integration into the system as a service. Thus this method allows the most diverse and complex quality assurance tasks to be executed quickly and easily – and without a single line of programming code. Figure 2: Ensuring the visual quality of the bubble beams, on the left to detect minimal damage to the surface, and on the right, the automation processes can be carried out efficiently and robustly. A ready-to-use prototype can be built in just a few hours and expanded into a productive solution within a very short period of time. Data Spree’s rapid AI models additionally ensure real-time capability in high-frequency production and logistics processes. Another benefit is the flexibility of the learning system. If packaging, packaging characteristics, or products change due to changes in production or logistics, AI can simply be “fed” with new images and retrained. This enables a fast and efficient response to changes in production or logistics without the need to start from scratch or purchase and implement a new solution. With Deep Learning DS, data from ongoing production processes and detected errors can be stored, managed, and statistically evaluated in the long term. In this way, the highest quality requirements can be continuously achieved by combining data management with AI training. Fast and easy implementation An individually trained AI model can be integrated into any client application via the standard ONNX open format. Data Spree Inference DS’s implementation environment also provides a simple graphical user interface through which an AI model can be quickly implemented on the respective devices, such as a smart camera or industrial computer, using the principle of drag and drop. This saves integration time – and on top of all that costs. ADLINK and Data Spree Partnership The ADLINK Edge ecosystem has built-in connectors for Data Spree’s Deep Learning DS. This provides the two companies with the ability to expand the capabilities of any Vision AI solution to ADLINK’s wide range of hardware and software in a scalable manner without changing the platform for both platforms. Using ADLINK’s broad IoT platforms, we can integrate with businesses in Edge that provide filtered data for enterprises and the cloud. By Manuel Haas and Chris Montague as co-founder of the startup Data Spree in Berlin, Manuel Haß implements a shared vision of automating the future: making deep learning accessible to everyone in order to automate cognitive processes. After studying computer science at TU Berlin and ABB terminals, Manuel worked on autonomous vehicles at DCAITI in Berlin before founding Data Spree. Chris Montague is ADLINK’s EMEA Head of Sales for Edge Solutions. An IoT expert with over 22 years experience in the hardware, software and IT solutions market, prior to ADLINK he worked for an IT consulting firm, advising clients and providing services from pre-sales to project delivery across multiple sectors. He earned a Bachelor’s degree in Computer Science from Northumbria University and began his career in information technology writing code to improve and simplify databases for large public sector clients. Chris Montague, Manuel Haas.