Localization of Barcodes Using Artificial Neural Network
Nikolay N. Ventsov, Подколзина Любовь Александровна · 2018
This article discusses a method for localizing barcodes using artificial neural networks (ANNs). The algorithmization of the presented method allowed writing a software with using Python language, which implements barcode recognition using the modified ANN. The goal of this work is the development of the barcode recognition algorithm based on ANN. Its allows to increase the accuracy of the barcode detection process on the images. The novelty of this work is to develop the structure of a convolutional neural network for algorithm of barcodes localization using artificial neural network. When preparing the training sample manually, the marking of the identified objects on a small number of images (about 800) was carried out. In connection with the relatively small number of barcode images, the data was augmented to 8000 images. To tune a neural network, a training sample is coded into input vectors of the neural network. A part of the encoding procedure is data normalization - mapping them to the interval [0, 1]. After conducting the training and determining typical errors, the authors gradually added various types of distortions to the existing training set of images, corresponding to the most common types of errors. The number of “distorted” images added was varied and selected according to the feedback of the validation sample. Authors developed the structure of a convolutional neural network for algorithm of barcodes localization. For source images, the share of correct barcode detection using this neural network was 97.41%. For rotated images, the percentage of correct detection was 85.68%, and for images with changes in brightness-83.32%. In comparison with similar approaches, the improvement was 1.32%, 7.21% and 10.68%, respectively.