Barcodes detection improvement via weakly labeled data

Dmitry A. Zvonarev

Moscow Institute of Physics and Technology

Weakly-supervised neural network object detection is used when a large amount of labeled data is not available. The results of experiments in various studies show that the quality of weakly-supervised models does not exceed the quality of fully-supervised models. Proposed approach improves the quality of barcodes detector and reduces the cost of obtaining markup, using a small amount of labeled data and a large amount of unlabeled data. The quality of the model trained on a small part (169 examples) of labeled data: Precision = 0.627, Recall = 0.869, F1 = 0.728, the quality of object classification regardless of objects type: Accuracy = 0.624. The quality of the improved model trained on artificially labeled data (2531 examples) and tuned on labeled data (169 examples): Precision = 0.856, Recall = 0.892, F1 = 0.874, quality of object classification regardless of objects type: Accuracy = 0.924.

convolutional neural network, barcode, weakly-supervised object localization, deep learning, object detection, self-supervised learning

Back