倒傳遞類神經網路表面瑕疵分類雲端系統

A Cloud System for Back-Propagation Neural Network based Surface Defect Classification

江承恩(Cheng-An Chiang)

指導教授 ： 章明

中原大學/工學院/機械工程研究所/碩士(2015年)

https://doi.org/10.6840/cycu201500835

摘要

為了滿足高速高精度自動化檢測與遠程監控之需求，勢必要有更快速更詳細的分類演算法並結合雲端儲存系統，故本研究目的在開發一套結合雲端儲存之瑕疵分類系統，瑕疵經過分類後將檢測結果儲存至雲端伺服器中，透過網頁呈現給遠端使用者。
具體作法係以Hu矩描述瑕疵輪廓特徵並結合特徵影像亮度參數，當特徵值經過正規化後，利用倒傳遞類神經網路(Back Propagation Neural Network，BPN)來做此特徵資料與瑕疵類別之訓練，分類結果則儲存至MySQL所建立的資料庫上，使用超文字預處理器(PHP: Hypertext Preprocessor，PHP)進行阿帕契(Apache)伺服器與資料庫的溝通，使用者就可透過網頁觀看檢測結果。
本研究以觸控面板作為測試樣本，以擷取350*350像素、每一個像素尺寸為3.5mm之影像進行訓練，經過1000次迭代後，倒傳遞類神經網路的均方差(Mean Square Error，MSE)為0.0019，取其權重w與偏權值，以刮痕、氣泡、粉塵瑕疵做分類測試，共20張影像，每張影像分類過程耗時低於0.0001秒，準確率約90% 。

關鍵字

倒傳遞類神經網路 ； 瑕疵分類器 ； 雲端系統

並列摘要

In order to meet the industry needs for high-speed high-precision automated detection and remote monitoring, a defective part that has undergone line inspection is subjected to a fast classification scheme consisting of an algorithm which can feed the results directly to a cloud storage server. The purpose of this study is to develop a combination of cloud storage and defect classification system where the detection results are stored to the cloud server and then accessed via web presentation to a remote user.
The classification of defect profile is implemented by means of comparing a defect’s luminance characteristics to a target pixel value using the Hu set of invariant moments. The characteristic values normalization and the Back-Propagation Neural Network (BPN) model are implemented to train the algorithm to recognize defect information their characteristics. The classification results are saved to an established Mysql database and use PHP to communicate with the database side of the page. The user can view and access the test results via a website.
Defect maps from the optical inspection of touch panel glass samples were used to evaluate experimentally the classification algorithm. Each map records a 350 * 350 pixel region of the glass sample, with each pixel equal to 3.5 m. Different iterations of the BPN were implemented to seek optimized classification by comparing the mean square error (MSE) between target output and defect profile of the test sample. Defect classification of microscopic scratches, bubbles, and dusts on 20 different defect maps yielded accuracy rates above 90%. Classification of each image was completed in 0.1 ms.

並列關鍵字

Back-propagation Neural Network ； Classification ； Cloud system

參考文獻

[1] Mei Chih Hu and I Shou Tsai, "Fabric inspection based on best wavelet packet bases," Textile Research Journal, vol. 70, pp. 662-670, 2000.

[2] Gang Wang and T.Warren Liao, "Automatic identification of different types of welding defects in radiographic images," Ndt & E International, vol. 35, pp. 519-528, 2002.

[4] Issam Abu Mahfouz, "Drilling wear detection and classification using vibration signals and artificial neural network," International Journal of Machine Tools and Manufacture, vol. 43, pp. 707-720, 2003.

[5] Meng Xin Li, Cheng Dong Wu, and Yong Yue, "Automatic visual inspection and classification based on rough sets and neural network," in International Conference on Machine Learning and Cybernetics, 2003, pp. 3095-3099.

[6] Soo Beom Park, Jae Won Lee, and Sang Kyoon Kim, "Content-based image classification using a neural network," Pattern Recognition Letters, vol. 25, pp. 287-300, 2004.