摘要
自然场景文本定位检测是文本识别的研究热点之一。EAST算法是目前自然场景文本定位检测算法较为出色的算法之一,在ICDAR2015数据集上,有着较高的准确率和召回率。但EAST算法仍存在着感受野不够大、长文本检测效果不佳的问题。因此本实验对EAST算法进行改进,通过改进EAST算法的结构,加入了ASPP网络,扩大感受野,加入了BLSTM神经网络,增强了文本之间的关联,提高文本定位效果。实验结果表明,该算法在ICDAR2015文本定位任务上的召回率为77.84%,精确率为86.24%,F-score为81.82%,优于经典EAST算法。
Text location and detection in natural scenes is one of the research hotspots of text recognition. East algorithm is one of the most excellent algorithms for text location and detection in natural scenes. It has high accuracy and recall rate in ICDAR2015 Dataset. However, the sensitivity field of EAST is not large enough and the effect of long text detection is not good. Therefore, this experiment improves the EAST algorithm by improving the structure of the EAST algorithm, adding the ASPP network, expanding the receptive field, adding the BLSTM neural network, enhancing the relevance between texts, and improving the text location effect. Experimental results show that the recall rate, precision rate and F-score of ICDAR2015 are 77.84%, 86.24% and 81.82% respectively, which are better than the classical EAST algorithm.
出处
《计算机科学与应用》
2021年第1期167-175,共9页
Computer Science and Application
