本研究是利用機器視覺的技術來檢測具一致性之方向性紋路的物體表面是否有瑕疵的存在。由於傳統對一致性紋路的處理是採人工目視判斷,瑕疵判讀標準不一,且缺乏客觀性,因此,本研究將藉機器視覺的檢測方法,提供一較具客觀的技術以作判讀。 本研究之表面瑕疵檢測技術,主要是應用於具方向性之一致性紋路,藉由傅立葉轉換(Fourier transform)可凸顯一致性紋路之週期性特徵的特性及反傅立葉轉換(Inverse Fourier transform)之影像還原技術,來進行方向性紋路之表面瑕疵檢測與分析。原始灰階影像之方向特徵在其相對應之傅立葉頻譜中會使高功率之頻率元素以直線方式分布,且與原始影像中之方向正交,因此本研究利用霍氏轉換(Hough transform)將二維之傅立葉頻譜轉換成一維之霍氏累加值直方圖,以便從直方圖上之峰點找到傅立葉頻譜之高功率強度之方向角度,如此即可將傅立葉頻譜上對應於方向性紋路之高功率強度的分布去除,再利用反傅立葉轉換,以便將原始影像中的一致性紋路去除,而只保留瑕疵部份影像。最後再利用統計管制界限法,把瑕疵部份與非瑕疵部份區分出來,即可將瑕疵部份凸顯出來。本研究之實驗樣本乃針對切削工件及布紋等紋路,且檢測效果皆相當良好。
The purpose of this research aims at the use of the machine vision for inspecting defects on surfaces with directional textures. Many surfaces of man-made objects in industry such as machined workparts and textiles can be considered as directional textures. The method of inspecting surface defects is based on two-dimensional Fourier transform (FT) of the surface image and the restoration technique of the inverse Fourier transform (IFT). The Fourier spectrum is ideally suited for describing the directionality of periodic line patterns in a gray-level image. The directional characters of an original image clearly correspond to high-power frequency components that are distributed along straight lines and are orthogonal to original directions in the Fourier spectrum. The lines associated with high-power frequency components in the power spectrum are detected by using the Hough transform. It requires only a simple one-dimensional accumulator to detect the peaks for all possible slope angles of lines in the Fourier spectrum. The frequency components falling on the detected lines or in the neighborhood of the lines are virtually eliminated by setting them to zero in the Fourier domain image. Then an IFT is applied to obtain space domain image. The IFT process will remove all homogeneous, periodic, directional textures in the original gray-level image, and preserves only abnormal features, i.e., defects, if they appear in the surface. Finally, the statistical process control principle is used to set up the control limit for distinguishing defects from noise in the IFT image. Experiments on real directional textures including machined surfaces, textiles and woods have shown promising results using the proposed approach.