Efforts toward Image Recognition Technology Using AI
- Methods for Reading Device Information -

Recognition of nameplates and management stickers of electricity meters using AI

Takada:TAKAOKA TOKO is working to improve operational efficiency and implement preventive maintenance using AI. It is my fourth year at the company. Since the second year, I have been involved in the research on the use of AI. This time, we tried to recognize product information by image-recognition using AI with the aim of management work efficiency of products (auto-recognition and bookkeeping of product information). The target of image recognition was our products such as electricity meters and product management stickers, and we used "machine learning" and its advanced form "deep learning" for AI.

Nakayama:"Machine learning" and "deep learning" make AI learn data, performing prediction and recognition based on that data. "Machine learning" is further divided into "supervised learning" and "unsupervised learning." "Supervised learning" is a method of creating a set of pre-collected data and answers to it and giving it to AI. When AI is given data, it learns to output the corresponding answer. On the other hand, "unsupervised learning" is a method for analyzing a large amount of data that does not have an answer and knowing the essential structure of the data, and is used for big data analysis. This time, we used "supervised learning."

Takada:Since "deep learning" abstracts data to a higher degree than "machine learning" and learns by extracting features, high versatility and recognition/prediction performance can be expected. However, a high level of data abstraction requires more data, which is also a weakness of "deep learning."
On the other hand, there are methods such as data padding that intentionally adds noise to the data, transfer learning that uses the results of learning with other data, and fine tuning. These methods are also used this time.

Nakayama:We have considered AI that detects the area where information is written from the nameplate of the power meter using "deep learning" and "machine learning." As a result, the "object detection algorithm" of "deep learning" achieved a reliability of 90% or more. We also considered an "object detection algorithm" based on "machine learning," but it was not possible to deal with minor differences such as the shooting angle and lighting. The advanced abstraction of data by "deep learning" has the feature of being able to absorb such differences, leading to the detection of highly reliable areas.

Takada:On the other hand, we came to the conclusion that we should combine "machine learning" and "deep learning" for management stickers. The QR code of management sticker is used as a "marker," and it is detected by the "marker detection algorithm" of "machine learning." By using the results, it is possible to limit the range of area detection by the "object detection algorithm" of "deep learning," so it was possible to reduce the false detection of the area due to the background.

Efforts to learn the unknown field of AI and connect it to new businesses

Nakayama:We don't have any employees who have been working with AI for many years, and AI itself is not a highly established technology that consistently produces definite results. Therefore, regardless of age, everyone in our group begins by learning about AI from scratch. We hold group-wide seminars on AI and participate in external seminars. We research products, technologies, papers, etc., related to AI that are announced daily. Each member involved in AI research is assigned a theme and is working on it. We report on the status of each theme about once a week and exchange opinions.

Takada:When we introduce AI into our company, there is also a vague sense of hope that "AI will solve problems." However, in order to successfully utilize AI and introduce it on a full-scale, it is necessary to clear various issues such as collecting the necessary data and setting appropriate goals. In order to convey such information, we are making efforts such as presenting the information at an in-house conference held about twice a year.

We are working with the idea that the introduction of AI will lead to the improvement of internal work efficiency and the reduction of routine work. People make a mistake, which is the same for repeated simple tasks. We aim to create an environment where AI can do the work that can be entrusted to AI, and humans can concentrate on more creative work that only humans can do.

Nakayama:There is a possibility that we may end up with a so-called PoC (Proof of Concept), where we introduce a little bit of AI as a trial. We have never had a long history of AI, but we have the advantage of being close to our in-house field and product developers. Taking the advantage of that, we aim to improve operational efficiency and apply it to products in cooperation with the company so that it does not stop with PoC. Furthermore, I would like to develop unique products and services that utilize AI and connect them to new businesses.