课程学习方法中文字识别算法研究

场景文字识别是一个极具挑战性的研究方向,有着重要的应用价值。但是由于文字表现形式丰富多样,识别结果大多不尽如人意。针对此问题,本文提出了基于课程学习的训练方法。该方法对数据集进行排序得到一个难度提升的训练序列,而不是随机地从数据集中选择训练样本,使得模型在训练初期能够学习到更加精确的特征,提高了模型的鲁棒性。通过实验分析,本文所提出的方法可以加快模型的收敛速度,使用不同课程序列训练ASTER算法在COCO-Text数据集上得到1.8%、1%的提升,CRNN算法在COCO-Text数据集上得到0.2%的提升。

Researches on Character Recognition Based on Curriculum Learning

Scene character recognition is a challenging research direction, which has important application value. However, due to the rich and diverse forms of characters, most of the recognition results are unsatisfactory. To solve this problem, this paper proposed a training method based on curriculum learning. This method sorts the data set to get a training sequence with increased difficulty instead of randomly selecting training samples from the data set so that the model can learn more accurate features in the early stage of training and improve the robustness of the model. Through the experimental analysis, the proposed method accelerates the convergence speed of the model, using different course sequences to train the ASTER algorithm to get 1.8% improvement on the COCO-Text data set, and the CRNN algorithm gets 0.2% improvement on the COCO-Text data set.