Image Recognition of Numbers by AI Convolutional Neural Networks

In recent years, AI technologies such as machine learning and deep learning have been rapidly advancing and beginning to be introduced into business due to improved CPU performance and big data.
AI is also expected to be used as a countermeasure for labor shortages and productivity improvement, which are social issues. For example, AI is being introduced in a wide range of fields, including unmanned cash register management in the retail industry, nursing care robot control in the medical field, pesticide spray management in the agriculture and forestry industry, and detection of abnormal products through image recognition in the manufacturing industry.

Generally, AI begins by providing educational material data and teacher data and allowing the system to learn. For example, a dog or a cat is shown as the teaching material data, and the answers that it is a dog or a cat are given as the teacher data. After successful learning, the AI will be able to infer that what it sees in the photo is a dog, even if it is shown a photo of a dog that is not part of the teaching data.

By introducing AI technology into meters, one of Aichi Clock Electric's core products, further value-added creation can be expected, such as improving the accuracy of flow measurement by removing noise superimposed on the measurement signal, and realizing remote automatic meter reading by recognizing images of numeral cars.

As described above, AI technology can be applied in a variety of ways. In this report, we will introduce a case study of image recognition technology using convolutional neural networks (hereafter referred to as CNN) for numeral recognition.

About CNN

In CNN, an operation called convolution, which also appears in the image processing field, is applied to the input image. This operation can be used to enhance image features. While conventional machine learning requires a human to specify features, this method can extract features automatically.

In order to make accurate predictions, training must first be performed to create a brain (AI model). Repeat the process of giving an image for training and the correct value for that image. The internal parameters used for convolution operations, etc., based on the discrepancy between the image predictions and the correct values, are optimized using gradient descent or error back-propagation methods. This will result in an AI model specialized for number recognition (i.e., skilled at extracting the features of numbers), if, for example, images of numbers are used as training data.

Researcher's Voice

■What challenges or findings have you encountered in image recognition of numbers by AI?

AI models are created using deep learning techniques, which are currently the most popular AI technologies.
The created AI model is used to obtain the required answer (e.g., recognition of numbers from camera image data) by inputting measured data, but since 100% correct answers cannot be obtained and the theoretical basis of how the answer is derived from the algorithm cannot be seen, we feel that the developed AI must be used in a black box state. Therefore, we feel that it is necessary to increase the reliability of the developed AI through repeated experiments and verifications.

Please let us know about further improvements and future refinements.

In this case, we created an AI model that recognizes and classifies numbers using a CNN, and mounted and operated the AI model on a microcontroller. In the future, we expect to improve quality and create more added value by incorporating AI into our main product, meters. To this end, we intend to accumulate more verifications by investigating and researching AI, the latest technology, to build a safer and more secure society.

Inquiries about research and development of the latest technologies