Anomaly Detection with Computer Vision
After a period of stagnation and deindustrialization, Europe is now in a process of change, betting on a strong and assertive green reindustrialization. This new vision of Europe gives rise to new challenges to deal with high productivity and a variety of products. Simultaneously, the new industrial approaches are based on small-scale production, with the creation of high technology tailor-made and customized products. This new concept of (re)industrialization is somehow typified in the Industry 4.0 paradigm. Globally, the main technical areas of this paradigm comprise items, such as artificial intelligence (AI), internet of things, machine learning, cloud systems, cybersecurity, adaptive robotics, and advanced automatic inspection. All these topics are promoting radical changes in business processes and organizations.
Automatic inspection systems that use computer vision are designed to assess the quality of raw materials/intermediate products/final products. In a broad perspective of the computer vision concept (CV), image processing and analysis algorithms should now incorporate the most advanced techniques of artificial intelligence. These new methods and techniques increasingly enable the computer to simulate human vision in quality assessment processes on production lines. Therefore, greater efficiencies and shorter cycle production times are obtained when compared with those achieved by a human operator.
In a large number of inspection quality systems, the conformity of the object is assessed by using a golden sample, which serves as a standard object for the analysis/detection of any anomaly in the object or in parts of it.
More and more multifaceted objects in terms of manufacturing processes and materials can lead to different visual cues. Such a trend of increasing complexity will demand a better definition of object conformity. Additionally, the non-conformities of the objects remain unknown until actually occur. Thus, in most current industrial applications, accessing abnormal objects could be a very challenging, or even impossible, task.
Moreover, the key issue to apply AI combined with CV to ensure quality control in industrial inspection is the huge amount of data that AI algorithms need to perform with high precision. It is noteworthy that a major issue in the industry is to provide the algorithms with adequate information for quality assessment, namely the necessary amount of good production parts (normal or positive samples) as well as the negative or abnormal samples, which means poorly manufactured/produced products/parts. So, fully supervised classification could be unrealistic when assuming the availability of huge labeled training data with both positive and negative samples.
Thus, to face all these limitations, including dealing with complex objects, a variety of processes, and the need for a large amount of positive and negative data, the research activity has been oriented to the development of anomaly detection methods. Depending on the production process the non-conformities can be very irregular, and thus, several types or classes of anomalies are necessary to characterize the universe of abnormal characteristics and apart those from the normal object class. So, a very interesting research & development issue is to reduce the false positives, keeping high accuracy detection rates.
In many real-world contexts, these new AI anomaly detection methods, based on deep learning, produce better results than conventional techniques. Such AI approaches include mainly unsupervised, semi-supervised, and weakly-supervised anomaly detection algorithms.
Neadvance has already developed several applications of AI combined with CV technologies for the inspection of objects and quality control of production processes. One of the achievements is focused on the automotive sector for analyzing color, texture and pattern drifts in plastic, fabric, metal pieces, and leather. Here, systems for the inspection of backstitch car seats, navigation units, tires, and rubber, among others have been developed. The main subject was concerned with the use of techniques and strategies to deal with three critical issues related to the lack of data: generalization; data imbalance; and optimization.
Following the path of sustainable technological growth, Neadvance will continue with its strategy of partnerships, namely within the scope of collaborative projects with several partners. A model example of these collaborations is the IntVIS4Insp – Intelligent & Flexible Computer Vision System for Automatic Inspection (POCI-01-0247-FEDER-042778) project, with Universidade do Minho and Centro de Computação Gráfica, for the development of AI techniques, such as, self-supervised classification; generative adversarial networks; one-shot learning based on siamese networks; ensembling classifiers and transfer learning.
Thus, it is Neadvance’s goal to develop an intelligent and robust system that can be applied to any industry that works with a wide range of data. In addition, the system will be performing tasks such as classification, detection, and even segmentation. Therefore, this new solution will open the capacity of any industry to ensure the high quality of its products, using state-of-the-art AI and CV techniques.
In short, AdVision will have a platform that will control the Agents of the production line in function of the images captured by the cameras of the vision systems. Thus, are introduced in the latter, allowing the management of the production line intelligently and in real-time.