基于EAST的文本检测算法改进研究

Research on An Improved Text Detection Algorithm Based on EAST

EAST具有较高的识别准确率,但上述算法受到感受野的限制,以及样本权重的不合理,导致其对长文本检测率较低。因此在AdvancedEAST的基础上对其网络结构进行改进,加入空洞卷积结构提高网络的感受野,增加1*5以及5*1的对称卷积结构以提升对横竖排长文本的检测效果。在ICPR2018数据集上的实验结果表明,对网络的改进,降低了算法的loss,提升了对长文本的检测效果,较传统EAST算法更优。

The EAST algorithm has a high recognition accuracy rate. However, the algorithm is limited by the receptive field and the unreasonable sample weights, which results in a low detection rate of long text. Therefore, on the basis of AdvancedEAST, the network structure is improved, and the dilated convolution structure is added to improve the receptive field of the network, and 1*5 and 5*1 symmetric convolution structures are added to improve the detection effect of horizontal and vertical long text. The experimental results on the ICPR2018 data set show that the improvement of the article on the network reduces the loss of the algorithm and improves the detection effect of long text, which is better than the traditional EAST algorithm.