小波轉換於自動表面瑕疵檢測與圖形比對的應用

Automatic surface inspection and pattern matching using wavelet transforms

蔣政輝(Cheng-Huei Chiang)

指導教授 ： 蔡篤銘 博士

元智大學/工程學院/工業工程與管理學系/碩士(2000年)

摘要

小波轉換﹙wavelet transform﹚的多尺度﹙multi-scale﹚、多解析﹙multiresolution﹚能力在許多影像處理的研究中已驗證，適合用作紋路分析與影像特徵擷取。就紋路分析方面，多解析能將影像分解成平滑子影像﹙smooth subimage﹚及細節子影像﹙detail subimage﹚，在檢測上可分離影像中重複性之規則紋路與區域性變動﹙local deviation﹚，而多尺度的特性經由解析階數的增加，分解出多個細節與平滑子影像，能更有效分離規則紋路與瑕疵；就影像特徵擷取方面，多解析能力所產生之細節子影像可擷取影像的細節特徵，而多尺度能力可縮小比對影像範圍，有利於圖形比對上的應用。基於小波轉換的特性，本研究將使用小波轉換技術於自動表面瑕疵檢測與圖形比對兩方面。
在自動檢測應用方面，本研究主要利用小波轉換與影像還原的技術於規則紋路的表面瑕疵檢測上，規則紋路分為結構式與統計式紋路兩類，利用無瑕疵之規則紋路影像作為標準，透過小波轉換之子影像能量值的分析，自動決定最佳的影像還原策略，包括還原用子影像及解析階數等，以使得規則紋路與瑕疵能有效分離。在圖形比對應用方面，則透過小波轉換的處理，將參考圖形與待測影像作多階解析而產生細節子影像，在細節子影像中以環形投影轉換針對高能量影像點作比對，不但可有效縮減影像比對範圍，還可避免逐點、逐角度比對上的耗時。

關鍵字

小波轉換 ； 影像還原 ； 瑕疵檢測 ； 圖形比對 ； 環形投影

並列摘要

Wavelet transform has been verified to be a powerful tool for texture analysis and acquisition of pattern features due to its multichannel and multiresolution characteristics. In this research we apply wavelet transform techniques for both defect detection and pattern matching applications.
In automatic surface inspection, we implement the forward wavelet transform to decompose an original image into smooth and detail subimages in different multiresolution levels, and restore specific subimages using the backward wavelet transform to separate defects from regular textured surfaces. The regular textures investigated in this research can be structural textures such as machined surfaces and statistical textures such as cast surfaces. By properly selecting the decomposed subimages and number of multiresolution levels, the restored image will remove repetitive texture patterns and retain only local anomalies. In this research, we have proposed a systematic procedure that can automatically select the best decomposed subimages and number of multiresolution levels for image restoration. The selection criteria and based on the energy responses of the wavelet coefficient matrix derived from the wavelet decomposition.
In pattern matching application, we decompose an original image into different resolution levels, and use only the high energy pixels in the detail subimages to compute the similarity (coefficient) between the model pattern and the scene image. This significantly reduces the computational burden of the traditional pixel-to-pixel matching. To make the matching invariant to rotation, we use a ring projection to represent object patterns. The ring-projection transforms the 2-D image in a circular window into a 1-D gray-level signal as a function of radius. Experiments on a variety of industrial objects have shown that the proposed methods are very efficient and effective for the applications of defect detection and pattern matching.

並列關鍵字

Wavelet transform ； Defect detection ； Pattern matching ； Textured surfaces ； Multiresolution ； Multichannel ； Ring projection ； Rotation-invariant

參考文獻

[1]D. Androutsos, K. N. Plataniotis and A. N. Venetsanopoulos, “Extraction of detailed image regions for content-based image retrieval”, Acoustics, Speech and Signal Processing, Vol. 6, pp. 3713-3716, 1998.

[2]W. Bin, Y. Motomura and A. Ono, “Texture segmentation algorithm using multichannel wavelet frame”, IEEE International Conference on Computational Cybernetics and Simulation, Vol. 3, pp. 2527-2532, 1997.

[3]V. N. Alan, “Texture classification using wavelet frame decompositions”, IEEE Conference on Signals, Systems & Computers, Vol. 1, pp. 311-314, 1997.

[5]K. O. Kim, I. S. Jung and Y. K. Yang, “High resolution image classification with features from wavelet frames”, IEEE International on Geoscience and Remote Sensing, Vol. 1, pp. 584-587, 1997.

[8]K. Hwang, S. Mandayam, S. S. Udpa, L. Udpa and W. Lord, “Application of wavelet basis function neural networks for NDE”, Midwest Symposium on Circuits & Systems. Part 3 (of 3), Vol. 3, pp. 1420-1423, 1996.