Back to Projects
AI Training Platform for Visual Inspection

AI Training Platform for Visual Inspection

A cloud-based, data-centric training platform with human-in-the-loop workflows for reproducible and continuously improvable visual inspection AI

Jun 2021 - Mar 2023 1 year, 9 months

Tech Stack

Data-Centric AIHuman-in-the-LoopML Training PipelineAWSDockerTensorFlowVue.jsFlask

Overview

産業用の外観検査 AI を対象に、
学習・評価・再学習をクラウド上で安定的に回すための AI 学習プラットフォーム のプロトタイプを設計・実装しました。

本プラットフォームでは、モデル改善を個別のチューニングに依存させるのではなく、
データとアノテーションを中心に、人の判断を組み込んだ Data-Centric / Human-in-the-Loop な改善プロセス を成立させることを重視しています。

※ 本ページは技術設計・アーキテクチャに焦点を当てたケーススタディです。
組織・製品・導入状況に関する詳細は意図的に省略しています。


Background / Problem

外観検査 AI の開発では、モデル精度以前に 学習プロセスとデータ管理がボトルネック になることが多くあります。

その結果、
モデルやハイパーパラメータを調整し続ける Model-Centric な試行錯誤 に陥り、
人の判断を活かしたデータ改善のループが回らない 状態になりがちでした。


Approach & Architecture

本プロジェクトでは、
Model-Centric な最適化から、Human-in-the-Loop を前提とした Data-Centric な改善ループへの転換 を設計の中心に据えました。

Design Principles


System Structure

Data & Annotation Layer

Human-in-the-Loop / Active Learning

Frontend

Backend API

ML Training Pipeline

Cloud Infrastructure


Role


Result / Impact


Lessons / Insights


Tech Stack

Infrastructure (Implementation Example)

View All Projects