基于DO-Conv的DBNet文本检测方法研究

Research on DBNet Text Detection Method Based on DO-Conv

目前,文本检测任务主要有基于回归的方法和基于分割的方法,基于回归的方法能够完成一般的文本检测任务,但是较难处理弯曲文本,基于分割的方法利用了图像分割,通过后处理的方式可以检测弯曲文本,得到较好的包围曲线,但是同时也增加了处理的步骤和预测的时间。DBNet方法提出可学习阈值并设计了一个二值化函数,简化了后处理的步骤,在文本检测任务中取得了很好的效果。随着越来越多的网络被提出并且在计算机视觉中都有不错的效果,针对基于分割方法的后处理复杂和预测速度慢问题,在DBNet方法上应用了较为新的ResNeSt网络,同时引入了DO-Conv卷积方式。实验结果表明,该方法在多个指标上都要优于DBNet方法,有较好的文本检测性能。

Currently,the main text detection task is based on regression method and the method based on segmentation,on the basis of regression method can complete the general text detection task,but more difficult to deal with curved text,used the image segmentation method based on segmentation,through post-processing can detect bend text,get bet-ter surrounding curve,but it also increases the number of processing steps and improves the prediction time.DBNet method proposed learnable threshold and designed a binarization function,which simplified the post-processing steps and achieved good results in text detection tasks.As more and more networks are proposed and have good effects in comput-er vision,this paper applies the new ResNeSt network in DBNet method and introduces the DO-Conv convolution method to solve the problems of complex post-processing and slow prediction speed based on segmentation method.