Neural Networks for Barcode Reading: a pilot study in the use of neural networks

Abstract

As one of the departments that helps NXP maintain its leading position in the semiconductor industry, ITEC develops, delivers and services equipment for state of the art high volume-low cost assembly solutions. ITEC is no stranger to the use of machine vision to enhance their production. To investigate the potential of neural networks the task was set to read barcodes with regular cameras, to replace the current hardware solution. In this thesis two different designs of neural networks are compared with a specially designed “heuristic” program. The three programs are tested for each program’s ability to decode 1D barcodes (the Code 3 of 9 symbology) and which program performs best when it comes to reliability/robustness, speed and cost. They will be tested for each of the three approaches. The test consists of computer generated image sets, based on the available original barcodes, as well as “realistic” image sets. The realistic images are synthesised barcodes that were printed and subsequently imaged using the camera set-up expected to be used in the production machines. Both these image sets contain horizontal shift, vertical shift and salt-and-pepper noise at three levels. Furthermore, the tests include real images of the original barcodes, taken with a camera. The results of the computer generated sets show that all three programs (both neural networks and the heuristic program) are successful in reading barcodes without noise in the images. The heuristic program completely fails to cope with noise, whereas the neural networks sustain performance. These results are mirrored in the results of the realistic sets. The proposed solution offers a promising alternative to the heuristic program when looking at the speed, reliability and costs. With a little further refinement it can even serve as a replacement for the current hardware solution.