Character Recognition Based on Multi-scale Feature Fusion Transfer Learning

摘要

Video surveillance has been using in daily life widely and has significant impact in security field. Reading time information in video has becoming an essential part. Because of the complex background of timestamp and the high similarity between targets, the features extracted by traditional machine learning algorithm cannot meet the requirements and cannot achieve the required detection speed and accuracy. In this paper, we improved the algorithm with multi-scale features. Features extracted from convolution layers of different scales are used to predict, which could improve the recognition accuracy of small targets. In addition, because the size of the collected dataset is 2860 and the amount of data is not large enough, we use data augmentation and transfer learning to train. Pre-training on the public dataset SVHN firstly, and then trains on the special dataset. Experiments show that the algorithm can recognize timestamps accurately.