RESEARCH 02
Object Detection & Industrial Vision
OBJECT DETECTION / INDUSTRIAL VISION
From quality control on manufacturing lines to agriculture and service robots — we solve real-world problems with AI that "sees and judges."
From quality control on manufacturing lines to agriculture and service robots — we solve real-world problems with AI that "sees and judges."
In manufacturing and infrastructure maintenance, defect inspection and equipment checks still rely heavily on manual labor, creating major challenges around labor shortages, inconsistent inspection quality, and cost. At the same time, missing a defect can lead to serious accidents or loss of trust.
Fast, high-precision visual recognition powered by AI is key to solving these problems. Industrial settings pose unique difficulties, however — extremely few defect samples, previously unseen anomalies, and the need for real-time performance. We research practical detection technologies suited to these real-world constraints.
Adapting YOLO-family models specifically for industrial imagery. With newly designed residual blocks and neck structures, we detect fine PCB defects at 93.4% mAP and 73 FPS, balancing accuracy and speed.
Researching methods that learn only from normal samples yet detect previously unseen defects. PatchCore, memory banks, Segment-Element-based Anomaly Detection (SEAD), and spatial/channel bidirectional attention together capture both structural and logical anomalies.
Systematically comparing methods that transfer knowledge from high-accuracy teacher models to lightweight student models, building anomaly detection models suited for edge deployment.
With Japan's shrinking workforce, labor-saving through robots and AI is a critical challenge. We are developing a system where a camera mounted on a mobile manipulator's arm recognizes dishes with AI, automatically collects them, and transports them to the kitchen. Demo video (YouTube) ↗ Download the dish-recognition app ↗
Early prediction of metal degradation is critical for industrial risk avoidance. We use machine learning to forecast degradation progress and support equipment maintenance planning.
Deployed across a wide range of fields, including fruit quality grading (over 99% accuracy), bridge fatigue-crack detection, and pest/disease detection in agriculture.