知识与财富的链接
针对现有优秀的anchor-free文本检测方法只挖掘了文本框几何特性而没有考虑文本框位置特性且缺乏有效的过滤机制,提出了挖掘文本框位置特性的anchor-free自然场景文本检测方法.该方法以ResNet50作为卷积神经网络的主干网络,将多个不同尺寸的特征层融合后预测文本框的几何特性和位置特性,最后辅之以二层过滤机制得到最终的检测文本框.在公开的数据集ICDAR2013和ICDAR2011上F值分别达到了0.870和0.861,证明了该方法的有效性.
The exiting excellent text detection methods only mine geometric feature of text boxes and lack an effective filtering mechanism. Aimed at this problem, this paper proposed an anchor-free text detection method through mining text box location feature. The method used ResNet 50 as the backbone of CNN(convolution neural network), and then fused the features with different size to predict the geometric feature and location feature of text boxes. Finally, this method used a two filtering mechanism to obtain the final detection text boxes. The method got the value of ◢F◣ 0.870 and 0.861 on ICDAR 2013 and ICDAR 2011 datasets, which demonstrate the effectiveness of mining text box location feature for anchor-free text detection in natural scene images.
