Toshiba Develops Image Recognition-Based Few-Shot Object Detection AI for Immediate Detection of a New Object with Unprecedented Accuracy by Registering a Single Image of the Object

-Ease of incorporation and operation contributes to digitization and promotes digital transformation at factories and plants that handle many different parts and equipment-

25 May, 2022
Toshiba Corporation


TOKYO—Toshiba Corporation (TOKYO: 6502) has developed Few-Shot Object Detection AI, an image recognition-based technology capable of immediately detecting a new object with unprecedented accuracy. Simply registering a single image primes the technology to recognize new products and parts in production and logistics. The company designed the new AI to be useful in situations where new objects appear frequently, such as plants that regularly process updated and improved products and parts, and distribution centers that are constantly required to handle new merchandise.
There are worldwide efforts to convert sensing data from the field into valuable information and return it to the field as feedback, thereby increasing productivity, quality, and efficiency in operations. Image recognition AI, which detects people and objects in images, is particularly valuable for analysis aiming to improve productivity, automate processes, and reduce workloads. However, one drawback of the technology is the need to prepare large amounts of image data for training to detect new objects every time they are needed in the situations where the technology is used.
Toshiba’s new Few-Shot Object Detection AI is automatically pre-trained to recognize the shapes of non-target objects in the training images, a proprietary development in the training process that enables the AI to detect new objects with the registration of a single image and no retraining in the field. In an assessment of detection accuracy based on a public dataset(*1) , the technology achieved unprecedented accuracy(*2) .
Companies that operate factories and plants that handle many different parts and equipment have been reluctant to incorporate AI into their processes because of the time and effort required for retraining every time something changes. Few-Shot Object Detection AI is easy to introduce and use, helping these companies digitize their operations and promote digital transformation to improve productivity, quality, and efficiency.
Toshiba is scheduled to give a presentation on the details of the new technology on May 25, 2022 at ICIAP 2021, the 21st International Conference on Image Analysis and Processing.

Development background

The AI business is expected to grow continuously, and the global AI market size is projected to be $340 billion by 2025(*3) . Image recognition AI in particular is coming into practical use in many sectors as a vital technology for recognizing the behavior of humans as well as their surroundings. Benefits include improved quality and productivity in manufacturing, more efficient field inspections of public infrastructure, and streamlined operations in distribution and logistics.
The Toshiba Group has used image recognition AI as part of its services based on ample knowledge and a long track record in the infrastructure sector, and frequently encounters in the field objects that did not exist when the company introduced the AI. For example, when handling new parts or products during long-term operations or adapting to a new plant that handles different parts or products, objects that have not yet been learned by the AI need to be added for analysis by the AI.
Normally, AI must be retrained in order to detect new objects. Retraining-type AI is difficult to use in situations where new objects appear frequently because the retraining requires considerable time and preparation of large amounts of image data and reference data in the field. In contrast, registration-type AI, which does not require retraining, does not detect objects accurately enough to be used in the field.

Features of the technology

To address these issues, Toshiba has developed Few-Shot Object Detection AI, which is capable of immediately detecting a new object by registering a single image of the object without retraining.
Conventional AI regards everything other than the target object in an image as background, and is trained on deep learning models that identify as proposal objects the areas in images where target objects are estimated. Toshiba developed a new method of automatic training that includes non-target objects treated as background (Figure 1). A pre-trained deep learning model was used along with a further developed method in order to establish Few-Shot Object Detection AI, which detects a new object from images by comparing proposal objects automatically identified from the parts typically identified as background to the newly registered object (Figure 2). Using this AI makes it possible to immediately detect a new object by registering a single image of it (Figure 3).
This new method achieved a detection accuracy score of 46.0%, the world’s highest mark for a registration-type AI(*2)  and a vast improvement over the score of 21.2% by a conventional registration-type AI that, like the company’s new AI, does not require retraining of deep learning models(*4) .

Figure 1: Comparison of Training (Conventional Retraining-Type AI and the Proposed Method)
Figure 2: Object Detection by Few-Shot Object Detection AI
Figure 3: Example of Differences in Recognition Accuracy (Conventional Registration-Type AI and Toshiba’s New AI)

Future developments

The Few-Shot Object Detection AI facilitates the introduction and use of image recognition AI in the field situations where companies have so far been reluctant to apply this technology. Toshiba designed the technology to promote digital transformation, thereby improving productivity, quality, and efficiency of operations in a wide range of the field situations.
The company intends to work toward a pilot project for the new technology as soon as possible, and to broadly apply it in the Group’s products and services with the aim of releasing it as a product in FY2023.

*1: Public dataset: The PASCAL Visual Object Classes Challenge

*2: Toshiba research as of February 2022
*4: Qi Fan, et al., Few-Shot Object Detection with Attention-RPN and Multi-Relation Detector, Conference on Computer Vision and Pattern Recognition, 2020.