Use case 實際案例 - 工廠光學瑕疵檢測 Optical Inspection

industrial_inspection.png

AI視覺結合邊緣運算-工業瑕疵檢測

AI Vision with edge computing - Industrial Defect Inspection

I. 企業面臨的挑戰 (The Challenge)

In the field of industrial manufacturing, defect inspection is a very important issue. For PCB manufacturer, wafer foundry, or LCD panel manufacturer, it is necessary to do defect inspection efficiently in their mass production procedure to get better yield rate and efficiency at the same time. The traditional AOI vendor develops AOI instrument to assist in the task. By adapting optical method and algorithm, they reached the goal in certain extent. However, there still exist some problems. The results of AOI have to be double-checked by operators because of the high false positive rate, which cause the efficiency reduction and the labor cost increasing. This challenge must be conquered if the enterprise wants to make more profit.

在工業生產的領域中,瑕疵檢測是一個非常重要的議題,不論是印刷電路板廠、晶圓代工廠、面板廠,這些大量製造的廠商為了提升良率,同時兼顧生產效率,必須有效率地做好瑕疵檢測。傳統的自動光學檢測廠商開發出了自動光學檢測儀器來輔助這項任務,透過光學與演算法,做到了一定程度的瑕疵檢測。然而,目前的自動光學檢測系統,實務上還是存在著問題,由於較高的假陽性率,自動光學檢測出來的結果,還是需要大量人力去做二次驗證,這種情況降低了產線的效率,增加了人力成本,因此企業若想提高獲利,必須去克服這個挑戰。

II. 解決方案 (The Solution)

Solution from AI4quant is based on three perspectives:

  1. Automatic AI vision: By training with massive data, the AI vision can obtain very high capability to distinguish false positive from positive report. Then the need of labor reduces gradually.
  2. Edge-computing: The model architecture with good balance between accuracy and efficiency is compatible with edge-computing. Therefore, the computing instruments used to do inference can be chosen considering cost-effectiveness.
  3. Self-adaptive model: With more and more labeling data, the ability of the model to distinguish will also increase. It’s like to accumulate resolving power of human into model, and it results in stable manufacturing efficiency that will not fluctuate with labor cost change.

AI4quant的解決方案包含下述三個層面:

  1. AI視覺自動辨識:藉由足夠資料的訓練,AI視覺辨識可以達到極高的辨假陽性的能力,因此能逐步降低人力的需求
  2. 邊緣運算:AI模型採用的是適用邊緣運算的深度學習模型,兼顧了準確度與運算效率,在選購設備上,能夠選擇符合成本效益的運算裝置做推論(Inference)
  3. 自我改善的模型:隨著標籤化的資料越來越多,模型辨識的能力也會越來越精準,等同於將人力辨識的能力累積起來,不會因為人力的替換而改變生產效率

III. 效益(The Benefits)

  1. Reduce the need and the cost of human resources.
  2. Increase manufacturing efficiency with faster recognition speed
  3. The solution take good balance between instrument cost and recognition speed, which can meet the requirement of industry.

  1. 降低了人力的需求,減少人力成本
  2. 更快的辨識速度可提高產線的效率
  3. 在設備成本與辨識速度上取得了很好的平衡,是能夠符合業界需求的解決方案

Use case 實際案例 - NLP

EC_dedupe_01.png

I. 企業面臨的挑戰 (The Challenge)

eCommerce companies such as Amazon or Walmart, provide cross-border online shopping for customers worldwide. It’s very attractive for consumers to have tens of millions of products from these platforms, some even with price comparison to help them find the best deals. Effective product matching (identifying duplicate products) is critical for user experience when it comes to online shopping. You don’t want your customers to search products from bunch of duplicate products, neither to give them irrelevant ones to be compared with what they really want.

電商公司像是亞馬遜或沃爾瑪,提供線上跨國的購物管道,它們不但擁有了數以千萬計的熱門產品,甚至還可且提供比價系統,為目標商品找出最好的價格,這些對於消費者非常具有吸引力。對線上購物來說,有效的產品配對 (product matching,辨識相同的商品) 對於線上購物的使用者體驗,是非常關鍵的:你不會想看到消費者在一堆重複的商品中挑選,也不會想提供不相關的商品讓他們進行比較。

However, sellers may have different descriptions for the same product. There are also different products with descriptions that are only different in attributes like color or size. For example, let’s see three titles of action figures as below:

  1. Marvel The Hulk, Multicoloured
  2. Marvel Avengers Age of Ultron Hulk Titan Hero Tech Action Figure
  3. The Avengers Marvel Titan Hero Series Hulk Buster Action Figure

然而,同一種產品,不同賣家可能有不同的描述;或者不同產品的描述差異卻只有尺寸或顏色等局部屬性的變化。我們用下面三種玩具公仔的主要說明 (title) 來做為例子:

  1. 漫威浩克,多種顏色
  2. 漫威復仇者之奧創紀元,浩克泰坦英雄科技公仔
  3. 復仇者漫威泰坦英雄系列:浩克毀滅者公仔

These action figures all have a keyword of Hulk, and it seems that the title is much similar between 2 and 3. In fact, 1 and 2 will be grouped together. For 2 and 3, they are two different products. It shows how difficult it is for product matching. For companies with thousands of millions of products and product is listed on and off every moment, it’s a challenge to have effective automatic-product-matching algorithm.

三個公仔都帶有Hulk (浩克) 的關鍵字,且2和3的文字敘述似乎比較相近,但實際上1和2才是類似的產品,2和3則是非常不同的產品。這顯示了一部分產品配對的困難。而對於一個具有千萬產品以上的電商公司,還要考慮商品上下架的動態變化,有效且自動化的產品配對 (automatic-product-matching) 是非常巨大的挑戰。

II. 解決方案 (The Solution)

Some eCommerce companies do product matching through universal product code (UPC) matching. Though it’s efficient, the accuracy is poor, products with different code have a chance to be the same, this will cause the customers to see lots of duplicate products. To address this issue, AI4quant builds AI model to effectively and efficiently find identical products for our customers.

一些電商公司通常藉由比對商品UPC等國際編碼來執行快速的產品配對,但是這依舊會產生許多錯誤的配對,但不同編碼產品依然有可能是同一件商品,這會使得客戶在瀏覽頁面時看到許多重複的商品,AI4quant為我們的客戶建構AI智能模型來有效辨識產品是否相同。

Solution from AI4quant is based on three perspectives:

  1. content similarity: AI model is built with deep learning, text mining and synthetic matching pairs to determine how similar a product description is to another one.
  2. image similarity: leveraging multiple online image databases with our customer’s data, we not only solve the issue of insufficient amount of labeled data, but also build a highly customized image recognition system
    (a)we also build image recognition system based on MobileNetV2, which provides excellent prediction accuracy and significantly reduces computation time, it can be used for industrial applications such as defect detection
  3. finally, an optimized product matching AI system was built based on content and image intellectual models mentioned before

AI4quant的解決方案包含下述三個層面:

  1. 內文相似性:藉由文本挖掘、合成配對樣本及深度學習建立AI辨識模型,評估一個商品敘述與其他敘述的相似程度
  2. 影像相似性:結合多種線上資料庫及客戶提供的資料,不僅解決客戶影像標籤 (image labeling) 不足的問題,更可建立出高度客製化的影像辨識模型
    (a)AI4quant的影像辨識AI可藉由MobileNetV2建構,其優點是在良好的預測準確度下,大幅降低運算量,可用於包含瑕疵偵測 (defect detection) 等工業量產應用
  3. 最後,一個結合前述文本與影像智能模型的最佳化產品配對AI系統產生出來

III. 效益(The Benefits)

  1. both sensitivity and specificity are more than 80%
  2. reduce computation time to 10% compared with traditional AI solutions
  3. increase customer’s satisfaction during online shopping (more revenue!)

  1. 靈敏度與特異性超過80%
  2. 運算量減少一個數量級 (約為傳統AI模型的1/10)
  3. 提高客戶線上購物時的滿意度 (更多的營收!)

IV. AI模型架構(AI model architecture)

EC_dedupe_02.png