Development of an Artificial Vision Algorithm for T-shirt Inspection

Abstract

The detection of defects in the textile industry is carried out by skilled labor to detect and classify the defects found in finished pieces. However, these personnel can make human errors that impair the quality of the final product. that is why three iterations of an artificial vision algorithm were developed for T-shirt inspection with the objective of providing the first step towards the automation of a process that has been manual for decades. This opens the door to continue automating manual operations in manufacturing workshops, to be able to implement artificial intelligence projects in the future in order to improve the efficiency of the processes and the quality of the inspections. To determine the functionality of the algorithms, experimental tests were carried out with 174 images in total. A first photo shoot was done for two colors of t-shirt: black and white; gray was also included to complete the samples and it was assumed that the slight color difference would not affect the performance of the algorithms. This first shot consisted of two batches named B1 and N1, after the initials of the colors. A second photo shoot was done in the same way with two batches named B2 and N2. Lastly, a third photo shoot was done, again with two batches called B3 and N3. An algorithm was developed for each photo shoot taken. The best method for the pictures was the 3rd one, consisting of a simple and solid background that didn't add noise to the image resulting from the algorithm. It was determined that, for white shirts, batch B2 with algorithm 2 obtained the best percentage of accuracy with a total correct detection of 31 out of the total sample of 40. For black shirts, batch N3 with algorithm 3 obtained the best percentage of accuracy with a total correct detection of 26 out of the total sample of 31.