本研究是利用機器視覺的技術來檢測結構性紋路與統計性紋路的表面是否具有瑕疵存在。由於1)小波轉換具有局部性(local)處理的能力,對於小區域之瑕疵能有效凸顯,2)小波具有頻率特性,使得在處理瑕疵上不易受環境影響,3)相對於頻率域之轉換方法,小波轉換處理速度快,因不須事先經過繁複的訓練與數學計算,使得小波轉換在速度處理上有較佳的效率,4)小波轉換不會變動影像物體的相對位置,且保留紋路與瑕疵的空間關係與影像大小,基於上述優點,本研究採用小波轉換技術偵測紋路表面之瑕疵。 傳統上,小波轉換應用於表面瑕疵檢測是擷取小波係數矩陣中紋路特徵值進行處理,由於特徵值選取不易,本研究將採用小波轉換之影像還原技術來檢測表面瑕疵,本研究方法乃利用小波轉換技術之影像轉換與影像還原兩大部份,藉由多尺度、多解析的觀念,配合本研究之還原平滑部份或細節部份之子影像兩策略來凸顯表面瑕疵,同時探討小波基底,還原階數與外在因素對瑕疵檢測之影響,實驗以結構性紋路與統計性紋路(包括布紋、沙紙、皮革、金手指及銑削和刨削工件)為測試樣本,由實驗結果得知紋路中瑕疵皆可有效的被檢測出來。
In this research we aim at automatic surface inspection using wavelet transforms for both structural and statistical textures. Small surface defects generally appear as local anomalies embedded in a homogeneous texture. The nature of the defects leads us toward the multi-scale and multi-resolution analysis method of wavelet transforms, which permits an efficient local spectral analysis. Wavelet transforms have been traditionally implemented for texture analysis and defect detection by selecting proper textural features in wavelet coefficient matrix. The proposed method does not reply on local texture features. It is based on an image restoration scheme using the wavelet transform. For each wavelet-transformed image, we obtain one smoothed residual sub-image and three detail sub-images, which contain fine structures with horizontal, vertical and diagonal orientation. By properly selecting the smoothed sub-image or detail images for backward wavelet transform, the restored image will remove homogeneous, periodic textures and signify only local anomalies. Experiment on structural textures such as machined surface and textile fabrics, and statistical textures such as sandpaper and leather have shown promising results using the proposed method.