Automatic detection of the texture defects of woven fabric on textile loom is important from both time and financial points of view. Since the types of defects do vary, depending on the source of fault, designing a general automatic defect detection system is very difficult. Various methods toward this problem have been developed and are being developed. Some of these methods are spatial, some are spectral and some others are feature based methods. In addition, some of methods take into account the color, texture and morphologic features of the fabric. Most of studies conducted in this concept utilize decomposition tools, filters, image features, co-occurrence matrices, and a decision system such as SVM or Artificial NNs. Despite of a massive work has been carried out in this direction, a satisfactory rate of success in this matter has not been reached yet. A successful method need to be generalized over variety types of fabrics as well as types of defect. Also for real-time applications the process (soft/hard) ware should be fast enough to eliminate motion effects. To avoid artifact or noise caused by vibration of the loom, fabric images should be captured at highest rate (fps or lps), and the resolution of camera (pixels/mm) need to be good enough high to capture the details of fabric. In this study, the fabric images were real time captured on the textile loom.The camera was a line camera with 25000 lines/sec. (lps) and 1024 pixels/line (BASLER camera). Totally 1493 (521 non-defected and 973 defected) images were processed through the designed system shown in Figure 5 (Şekil 6). First, two dimensional spectrums were obtained normalized and high pass filtered to remove the DC effect.With the assumption that the defect spectrum should be concentrate between DC and the dominant peak or component of the image, the images were filtered with a spatial pass-band kernel having non symmetric transition regions and hence cut-off spatial frequencies in x and y directions. Here, the kernel’s bandwidths 𝜎𝜎𝑥𝑥 and 𝜎𝜎𝑦𝑦 are chosen to be unequal (𝜎𝜎𝑥𝑥 = 25 and 𝜎𝜎𝑦𝑦 = 7.5 for pass bands in x and y directions and 𝜎𝜎 = 2.25 for stop band in both directions). By definition, this asymmetric property is coming from the texture of fabric. If the fabric texture would be symmetric then the cut-off spatial frequencies would be equal. The average power values calculated over these localized spatial spectra were found to be a significant feature for discrimination. The Euclidian distance of these power values from the one that calculated from the reference image (power of average of 10 non defective images) was then used as the classifying criterion. Among 1493 woven fabric images taken under process only 12 images were false alarmed by the designed system which is served a success rate about 99.12%. In addition, the developed system was successful and is fast enough to operate at real time basis. In studies conducted in this particular field the utilized spatial image features for evaluation were mainly obtained in terms of histograms and gray level co-occurrence matrices (GLCM), Gabor wavelet net (GWN), local binary pattern (LBP). In average the success rate from the processes motivated in this ways has not exceeded 96%. Also the process employing such methods are quite time consuming, which may not be appropriate for realtime applications. Therefore, as the designed system compared to the ones found in the literature, the superiority of the designed system is obvious, particularly in being a real time process, which is a highly desirable feature in textile industry.