基于EAST改进的任意方向场景文本检测

高效和准确的场景文本(efficient and accuracy scene text,EAST)检测算法速度快且结构简单,但是由于文本结构的特殊性,导致在检测中尺寸较小的文本会被遗漏,而较长的文本则完整性较差.针对EAST算法存在的问题提出一种新的自然场景文本检测模型.该方法利用自动架构搜索的特征金字塔网络(neural architecture search feature pyramid network,NAS-FPN)设计搜索空间,覆盖所有可能的跨尺度连接提取自然场景图像特征.针对输出层进行修改,一方面通过广义交并比(generalized intersection over union,GIOU)作为指标提升边界框的回归效果;另一方面通过对损失函数进行修改解决类别失衡问题.输出场景图像中任意方向的文本区域检测框.该方法在ICDAR2013和ICDAR2015数据集上都取得了较好的检测结果,与其他文本检测方法相比,检测效果也得到了明显提升.

Improved text detection in arbitrary directions based on EAST

The efficient and accuracy scene text (EAST) detection algorithm is fast and simple in structure, but due to the special nature of text information, smaller texts will be missed in the detection, and longer texts will be less complete. A new natural scene text detection model is proposed to solve the problems of EAST algorithm. This method uses the neural architecture search feature pyramid network (NAS-FPN) design search space to cover all possible cross-scale connections to extract natural scene image features. Then it is modified for the output layer. On the one hand, the generalized intersection over union (GIOU) is used as an indicator to improve the regression effect of the bounding box. On the other hand, by modifying the loss function, the algorithm can solve the category imbalance problem. Finally, the text area detection box is output in any direction. This method has achieved good detection results on both ICDAR2013 and ICDAR2015 datasets, and the detection effect has been significantly improved compared with other text detection methods.