『您的瀏覽器不支援JavaScript功能,若網頁功能無法正常使用時,請開啟瀏覽器JavaScript狀態』

跳到主要內容區塊

工業技術研究院

:::

技術名稱: 深度學習瑕疵分類技術

技術簡介

深度類神經網路之相關研究,近年來已獲得許多重要的突破,使得電腦視覺的物件分類應用範圍不再受限於人工設計物件特徵的瓶頸,而得以快速地擴展至不同的分類問題上。 本技術採用自主研發多分支出口之深層融合網絡架構(已申請專利),可支援不同大小之影像,確保準確度外,亦有效提升辨識速度;並且在無需重新訓練模型的情況下,於模型推論時可針對特定的類別抑制其漏檢率,提昇模型佈署彈性。此外,巨資中心研發之高效率標記資料篩選技術,可使用模型之識別能力狀況進行分析,在不需人力介入下有效找出少量對模型學習較有幫助之樣本,進而大量減少需人工標記之樣本,達到接近之模型訓練成效,以減少標記之人力成本及時間,使技術能快速導入產線。應用在諸多不同領域如警政安全、工廠監控、機器人視覺、自動光學檢測之瑕疵分類等,皆能有很好之效果。 此技術已被國內半導體製造、半導體檢測、PCB產業、PCBA產業…等各大業者採用,執行結果深受業界肯定。

Abstract

Deep Learning related researches have led to many substantial breakthroughs recently, making the application performance of image classification no longer be limited by the bottleneck of handcrafted features, and thus can be quickly spread to deal with different kinds of classification problems. The “Deep Learning Technologies for Defect Classification” of ITRI CITC adopted the patented Deeply-Fused Branchy Network architecture to take both accuracy and inference time into account to fulfil industrial strict requirements. In addition, given an already trained model, the miss rate of the class of interest can be further suppressed without the need of re-training model. Such flexibility greatly facilitates the production deployment. Applications in many different fields like security, surveillance, machine vision, and AOI, etc. can benefit from the technologies a lot. The technologies have been transferred to domestic companies, including the manufacturing industry of semiconductor, wafer inspection, PCB, and PCBA, etc. Empirical evidence confirmed the effectiveness of the aforementioned solution.

技術規格

.支援不同大小之影像 .2類分類可達99%以上、6類瑕疵分類正確率可達98%以上 (視影像資料之困難度有所調整) .可降低75%以上之假警報 .平均單張影像辨識所需時間少於20ms (使用中階以上GPU, 批次數量 > 8) .可提供Windows .NET、Windows DLL與Linux等平台之模組,系統整合簡單方便

Technical Specification

.Support images with variable resolutions. .Achieve 99% accuracy for binary classification, and 98% accuracy for 6-way classification (depending on the complexity of the image dataset). .Reduce the relative amount of false alarms more than 75%. .Averaged inference time per image is less than 20 ms (medium-class or above GPU, mini-batch size > 8). .Support Windows .NET, Windows DLL and Linux OS, greatly facilitate system integration.

技術特色

.本技術採用自主研發之多分支出口之網絡架構,可有效提升辨識速度及正確度,以滿足業界落地標準 .多類別瑕疵之預測最大化/最小化融合方法,可根據應用需求,提高特定類別的辨識正確度 .電腦自動選擇出具代表性之資料,再由人標記,協助克服初期標記資料過多之困難,並可大幅減少資料標記數量,加速深度學習技術之產業導入及應用

應用範圍

.自動光學檢測、機器人視覺、文件辨識、警政監控視訊分析、工廠人員監控應用等 .物件偵測分類核心技術之相關產業,皆可應用本技術以提升產業價值

接受技術者具備基礎建議(設備)

軟體:Windows 7 (含)以上、Windows Server 2012 (含)以上或Linux Ubuntu 14.04 (含)以上之作業系統。 硬體:含NVIDIA 1080 GPU以上之個人電腦或伺服器。

接受技術者具備基礎建議(專業)

視訊分析及熟悉軟體操作與資料分析基本概念。

技術分類 製造系統智慧化

聯絡資訊

聯絡人:李莉俁 執行長室

電話:+886-3-5916119 或 Email:villian.lee@itri.org.tw

客服專線:+886-800-45-8899

傳真:+886-3-5910257

[{"text":"企業網","weight":13.0},{"text":"材化所","weight":11.5},{"text":"機械所","weight":10.0},{"text":"綠能所","weight":9.4},{"text":"生醫所","weight":8.0},{"text":"半導體","weight":6.2},{"text":"南分院","weight":5.0},{"text":"太陽能","weight":5.0},{"text":"課程","weight":5.0},{"text":"遠紅外線","weight":5.0},{"text":"雷射","weight":4.0},{"text":"LED","weight":4.0},{"text":"LED可見光","weight":3.0},{"text":"5G","weight":3.0},{"text":"工研人","weight":3.0},{"text":"電光所","weight":3.0},{"text":"綠能與環境研究所","weight":3.0},{"text":"機械","weight":3.0},{"text":"資通所","weight":2.0},{"text":"面板","weight":2.0},{"text":"文字轉語音","weight":2.0},{"text":"實習","weight":2.0},{"text":"無人機","weight":2.0},{"text":"生醫","weight":2.0},{"text":"3D","weight":2.0},{"text":"v2x","weight":2.0},{"text":"員工","weight":2.0},{"text":"地圖","weight":2.0},{"text":"太陽光電","weight":2.0},{"text":"材料與化工研究所","weight":1.0}]