基于特征融合网络的自然场景文本检测

Scene Text Detection Based on Feature Fusion Network

目前,基于深度学习的自然场景文本检测在复杂的背景下取得很好的效果,但难以准确检测到小尺度文本.本文针对此问题提出了一种基于特征融合的深度神经网络,该网络将传统深度神经网络中的高层特征与低层特征相融合,构建一种高级语义的神经网络.特征融合网络利用网络高层的强语义信息来提高网络的整体性能,并通过多个输出层直接预测不同尺度的文本.在ICDAR2011和ICDAR2013数据集上的实验表明,本文的方法对于小尺度的文本,定位效果显著.同时,本文所提的方法在自然场景文本检测中具有较高的定位准确性和鲁棒性, F值在两个数据集上均达到0.83.

At present, scene text detection based on deep learning has achieved good performance in complex background.However, it is difficult to precisely detect text with small scale. To solve this problem, this study proposes a deep neural network based on feature fusion, and a new neural network with senior semantic is constructed by combining the highlevel feature and low-level feature of traditional deep neural network. Strong semantic information of the high layer network is utilized to improve the overall performance of the neural network, and the feature fusion network directly predicts text with multiple scales through multiple output layers. Experimental results on ICDAR2011 and ICDAR2013 datasets show that proposed method is significantly effective in detecting small scale text. Meanwhile, the proposed method has high accuracy and robustness in scene text detection, and the F-measure achieves 0.83 on both datasets.