Unsupervised Barcode Image Reconstruction Based on Knowledge Distillation

Abstract

Due to the influence of the lighting and the focal length of the camera, the barcode images collected are degraded with low contrast, blur and insufficient resolution, which affects the barcode recognition. To solve the above problems, this paper proposes an unsupervised low-quality barcode image reconstruction method based on knowledge distillation by combining traditional image processing and deep learning technology. The method includes both teacher and student network, in the teachers' network, the first to use the traditional algorithm to enhance the visibility of the barcode image and edge information, and then the method of using migration study, using the barcode image super-resolution network training to blur and super resolution, the final barcode image reconstruction using the depth image prior to in addition to the noise in the image; In order to meet the real-time requirements of model deployment, the student network chooses a lightweight super-resolution network to learn the mapping between the input low quality barcode image and the output high quality barcode image of the teacher network. Experiment shows the proposed algorithm effectively improves the quality and the recognition rate of barcode image, under the premise of ensuring real-time performance.