Adakotec, an anomaly detection AI company, has acquired an international patent for an anomaly detection algorithm that “detects anomalies with a small amount of data and visualizes the reason for judgment”
■ About Adacotech Adacotech is based on image analysis technology using the "HLAC feature extraction method" developed by the National Institute of Advanced Industrial Science and Technology (AIST). A technology company that provides visual inspection software that enables more efficient anomaly detection than before. As a track record so far, we are promoting the practical use of technology mainly in the automobile industry, such as introducing it to the actual production line of several companies, including major automobile parts manufacturers. ■Patent overview Patent number: 6903142 Name of invention: Abnormality detection device, abnormality detection method and program Date of patent acquisition: June 24, 2021 ■Patent content This patent was obtained for a small number of images of defective products. This is a technology that makes effective use of the With conventional anomaly detection technology, defective parts are illustrated in the form of a "heat map", but it is difficult to understand the reason for judging anomalies (defective products) and false alarms (good products but abnormalities are detected excessively). I had a problem. In addition to preprocessed images and heat map information, Adacotech's technology visualizes feature trends on a two-dimensional plane (anomaly discrimination plane).
This will increase the explainability and make it possible to achieve more accurate inspections.
The difference between "defect" (illustrated in red) and "false alarm due to misalignment" (illustrated in blue), which cannot be understood only by the heat map, can be interpreted by the "abnormality discrimination plane".
■Future development This time, patents were obtained in Japan and the United States, but similar patents are currently pending in Europe, China, and India. Adacotech aims to obtain patents in each country, and will use this patent acquisition as an opportunity to further accelerate its introduction to actual production lines in the domestic and overseas manufacturing industry. By spreading the automation of inspections and inspections, we will contribute to the improvement of productivity at manufacturing sites with technology, and we will promote the realization of our vision of "supporting the evolution and innovation of manufacturing." *1 HLAC feature extraction method: A versatile and high-speed feature extraction method with excellent recognition accuracy used for image analysis and recognition. When calculating the shape and size of an object to be inspected, in contrast to deep learning technology that performs complex processing, it is calculated only by sum-of-products operations of pixel values (values representing the color shade and brightness of each pixel). Because it is possible, it can be calculated instantly on a commercially available PC. In addition, due to the properties of position invariance (the recognition target can be recognized as the same even if its position changes) and additivity (when there are two targets, the sum of their features is the overall feature). , It is favorable for image recognition because it is not necessary to find the boundaries of regions that represent the same thing in the image (segmentation-free), and even if multiple abnormalities occur in the image, the features can be recognized individually. It is a feature extraction method with properties. Company name: Adakotec Co., Ltd. Established: March 2012 Location: 3-28-5 axle Ochanomizu, Kanda Ogawamachi, Chiyoda-ku, Tokyo Representative: Ryota Kawamura Capital: 100 million yen (not including capital reserve) Service Contents: Automation of product inspection and inspection URL: https://adacotech.co.jp/service