We are conducting research and development on natural language processing for chatbots. Along with our work on text classification and text similarity assessment algorithms used in systems such as chat support on our corporate website, we are also devising new technologies such as improving the quality of training data and the development of simple chatbots from small quantities of data.

We are conducting research and development on an AI OCR engine that applies deep learning and other similar techniques to key topics such as training on small amounts of real-world data, robustness for dealing with noise such as outdoor lighting, and handwriting recognition. The technology is used in a variety of in-house and external applications, including for scanning licenses, invoices, or other documents and for the automatic reading of freight container IDs. Patent applications have been made for some of these technologies; our unique work has delivered impresssive results.

Our extensive research and development includes the use of edge devices for image analysis, especially object recognition based on the analysis of video from fixed cameras. Technologies being developed for people recognition include tracking people's movements and detecting when they gather in the same place, following specific individuals across different cameras, and aligning camera recognition results with floor maps or other plans. We are also working on other related technologies, such as AI processing acceleration for edge devices and techniques for monitoring traffic conditions that include the estimation of vehicle speed and anomaly detection.

STAION is an AI image analysis platform that uses AI to analyze video acquired by cameras and convert it into the desired forms of information. In addition to communication networks and edge devices for the analysis and processing of camera video, we are also integrating STAION into services that use the results of AI analysis for age and gender identification, head counts, and detection applications such as out-of-stock products, empty seats, dangerous behavior, or quality defects. By doing so, we can serve as a one-stop source for video analysis solutions suitable for a wide range of industries and sectors.

We are working on the simulation of fish schooling to help develop feeding simulations that meet the needs of aquaculture. While the way in which fish feeding is done is extremely important for both the conservation of marine environments and for making production more efficient, the current practice is to have an experienced worker observe the behavior of the fish from above the waterline and decide based on intuition and experience. This work being done by SoftBank uses deep reinforcement learning to determine the different behaviors of fish when separated and when close together, comparing this with actual video for verification. The goal is to determine the feeding practices that best suit the conditions at the fish farm, including its size, the number of fish, and the fish species.

We have launched a quality standardization project that seeks to increase the value of fish by finding ways to measure freshness and taste. This includes building a machine learning model for characterizing the components that influence taste and use of a portable spectroscopy sensor for determining freshness in realtime. This work on the standardization of seafood quality also includes clarifying standards for frozen fish.

A joint research project with the Faculty of Fisheries Sciences at Hokkaido University on the use of AI in smart aquaculture for sturgeon has been running since February 2020. Experts have for some time been seeking to develop fish farming practices suitable for sturgeon, so far without success. Given that it takes the species at least six years to start producing roe and that they cannot be sexed until two to three years of age, a means of early identification is crucial to reducing both cost and risk. The goal of the project is to achieve efficient fish farming by using AI to recognize individuals automatically from camera video, allowing for early detection and the provision of care to fish laden with roe or exhibiting abnormal behavior.

We are investigating the analysis of posture in sports using AI video analysis, motion matching, and sports science. By developing functions for identifying how posture varies between athletes, analyzing where they can improve, and providing advice on how to do so, our goal is to be able to score athletes on ideal posture and deliver AI-based sports coaching.

This research seeks to overcome urban challenges by leveraging technology for digital twins and data on things like the movement of people, traffic, purchasing behavior, and visitors. In partnership with the University of Tokyo and Odakyu Electric Railway, it has taken the form of a two-year demonstration project in the Ebina Station neighborhood that commenced in April 2021. This has involved providing condominium residents, visitors to commercial facilities, and others with data-derived information such as favorable shopping opportunities or predictions of how busy eating or drinking establishments will be. It also provides commercial tenants with foot traffic predictions. As this helps them to stock appropriate products and manage their air conditioning, the benefits include reducing food loss and saving electricity. By utilizing these practices in other cities, SoftBank intends to create new value for communities.

Sakimiru is a demand prediction service that uses data on the weather and the movement of people. It predicts the number of customers and demand using an AI algorithm that was developed from weather data held by the Japan Weather Association together with statistical data on the movement of 120 million people extrapolated from the location information our base stations acquire from tens of millions of handsets. These highly accurate AI predictions facilitate the optimization of things like ordering food ingredients and shift scheduling.