Another technique that has been successfully tested for low cost implementation is based on some simple statistical information of the images captured. The process runs on two simultaneous tracks. The first track includes the calculation of mean and standard deviation block by block. Then the difference between the obtained values and the reference values are computed. A simple thresholding determines which blocks are defect-prone. A noise removal is done based on 4-neighborhood. In the second track, we first convert the image to a binary image to emphasize the structure of the fabric. Then in every block, the pixels are summed in the horizontal/vertical direction. The standard deviation of the sums is used to detect structural changes. If the value differs from the reference value by an amount bigger than a threshold, the block is considered defect-prone. A noise removal is done based on 4-neighborhood similar to the first track. This method has been suggested by Zhang et al., but was repeated here for making comparisons with other methods developed during my research.
 

Back to research page