Abstract
This research investigates the use of deep convolutional neural networks for racing bib number recognition in sport images. As labelled images of racing bib numbers are scarce, this paper investigates the potential benefits of transfer learning, by training models partly and fully on the Street View House Numbers (SVHN) dataset. Several deep neural network architectures are studied. The best recognition results were obtained by a model that had been trained on a hybrid dataset of the SVHN images plus an additional 262,131 images of racing bib numbers, with a recall of 0.93, precision of 0.95, and F-measure of 0.94 on the RBNR Dataset and a recall of 0.98, precision of 0.98, and F-measure of 0.98 on the private dataset. Our results are much higher than those reported in related work and show that deep learning can effectively be used for racing bib number recognition. Also, the effectiveness of transfer learning for this problem is demonstrated.
Recommended Citation
Ivarsson, Erica and Mueller, Roland M., "Racing Bib Number Recognition Using Deep Learning" (2019). AMCIS 2019 Proceedings. 20.
https://aisel.aisnet.org/amcis2019/data_science_analytics_for_decision_support/data_science_analytics_for_decision_support/20
Racing Bib Number Recognition Using Deep Learning
This research investigates the use of deep convolutional neural networks for racing bib number recognition in sport images. As labelled images of racing bib numbers are scarce, this paper investigates the potential benefits of transfer learning, by training models partly and fully on the Street View House Numbers (SVHN) dataset. Several deep neural network architectures are studied. The best recognition results were obtained by a model that had been trained on a hybrid dataset of the SVHN images plus an additional 262,131 images of racing bib numbers, with a recall of 0.93, precision of 0.95, and F-measure of 0.94 on the RBNR Dataset and a recall of 0.98, precision of 0.98, and F-measure of 0.98 on the private dataset. Our results are much higher than those reported in related work and show that deep learning can effectively be used for racing bib number recognition. Also, the effectiveness of transfer learning for this problem is demonstrated.