Abstract: At present, text detection algorithm under natural scene is a difficult challenge in the field of image processing. EAST algorithm is one of the natural scene text detection algorithms with excellent performance in recent years, which has a high recall rate and recognition rate, but there are still problems such as insufficient sensing field and unreasonable sample weight. Therefore, EAST algorithm is improved in this paper. The EAST network structure is improved and ASPP network is added to improve the sensing field. Loss is improved to optimize the problem of unreasonable sample weight, so that it has better detection effect on text localization. Experimental results show that while maintaining 18 frame per second, the proposed algorithm has a recall rate of 78.43%, accuracy rate of 85.78% and F-score of 81.94% in ICDAR 2015 text positioning task, which is better than the classical EAST algorithm.

Key words: text location, EAST algorithm, ASPP network, instance-balanced

摘要： 目前在图像处理领域，自然场景下的文本定位算法是一项具有困难的挑战，EAST算法是近年来性能比较出色的自然场景文本定位算法之一，具有较高的召回率和识别率，但是仍存在感受野不够大，样本权重不合理的问题。因此对EAST算法进行改进，对EAST网络结构进行改进，加入ASPP网络，提高了感受野，对loss进行改进，优化了样本权重不合理的问题，提高了对文本的定位效果。实验结果表明，提出的算法在保持18 f/s的同时，在ICDAR 2015文本定位任务的召回率为78.43%，准确率为85.78%，F-score为81.94%，优于经典EAST算法。

关键词: 文本定位, EAST算法, ASPP网络, 样本平衡