基于深度学习的场景文本检测算法研究

针对自然场景中任意形状文本图像因文本行难以区分导致的信息丢失问题,提出了 一种基于深度学习的场景文本检测算法。首先构建特征提取模块,使用Resnet50作为骨干 网络,在增加跨层连接的金字塔网络结构中引入并联的空洞卷积模块,以提取更多语义信息； 其次,对得到的特征图进行多尺度特征融合,学习不同尺度的特征；最后预测出不同内核大 小的文本实例,并通过尺度扩展逐渐扩大文本行区域,直到得到最终的检测结果。实验结果 表明,该方法在SCUT-CTW1500弯曲文本数据集上的准确率、召回率及F1值分别达到88.5%、 77.0%和81.3%,相比其他基于分割的算法,该算 法对弯曲文本的检测效果良好,具有一定的 应用价值。

Research on scene text detection algorithm based on de ep learning

To solve the problem of information loss of arbitrary shape text detect ion in natural scenes,a scene text detection algorithm based on deep learning is propo sed.Firstly,a feature extraction module was constructed,Resnet50was used as the backbone net work,and a parallel void convolution module was introduced into the pyramid network struc ture with cross-layer connections to extract more semantic information.Secondly,multi- scale feature fusion is performed on the obtained feature map to learn features of different s cales.Finally, the kernel size of text line of different scales is predicted and the predicted text line area is gradually expanded until the final detection result is obtained.Experimental re sults show that the evaluation indexes of accuracy,recall and F1-measure of the model on SCUT -CTW1500curved text dataset reach 88.5%,77.0% and 81.3%,respectively.Compared with othe r segmentation-based algorithms,this algorithm has a good detection effect on cu rved text;therefore,it has certain application value.