基于改进U⁃Net的低质量文本图像二值化

Degraded Document Image Binarization Based on Improved U-Net

文本图像二值化是光学字符识别的关键步骤,但低质量文本图像背景噪声复杂,且图像全局上下文信息以及深层抽象信息难以获取,使得最终的二值化结果中文字区域分割不精确、文字的形状和轮廓等特征表达不足,从而导致二值化效果不佳。为此,提出一种基于改进U-Net网络的低质量文本图像二值化方法。采用适合小数据集的分割网络U-Net作为骨干模型,选择预训练的VGG16作为U-Net的编码器以提升模型的特征提取能力。通过融合轻量级全局上下文块的U-Net瓶颈层实现特征图的全局上下文建模。在U-Net解码器的各上采样块中融合残差跳跃连接,以提升模型的特征还原能力。从上述编码器、瓶颈层和解码器3个方面分别对U-Net进行改进,从而实现更精确的文本图像二值化。在DIBCO 2016—2018数据集上的实验结果表明,相较Otsu、Sauvola等方法,该方法能够实现更好的去噪效果,其二值化结果中保留了更多的细节特征,文字的形状和轮廓更精确、清晰。

Text image binarization is a key step in Optical Character Recognition(OCR). However,the complex background noise of a degraded text image makes the extraction of its global context information and deep abstract information difficult.This results in an inaccurate segmentation of the text region and insufficient expression of features,such as text shape and contour,because of a poor binarization effect. Therefore,this paper proposes a binarization method for degraded text images based on an improved U-Net network.The segmentation network U-Net is suitable as a backbone model for small datasets,and a pretrained VGG16 is selected as the U-Net encoder to improve the model’s feature extraction ability.The U-Net bottleneck layer of a lightweight global context block is fused to realize the global context modeling of feature graphs. The residual skip connection is fused in each upper sampling block of the U-Net decoder to improve the model’s feature restoration ability. U-Net’s improvement is based on the aforementioned three aspects:encoder,bottleneck layer,and decoder,to realize a more accurate text image binarization. Experimental results on the DIBCO 2016-2018 datasets show that the proposed method can achieve a better denoising effect,retain more detailed features in the binarization results,and provide a better delineation of characters in terms of accuracy and clarity than Otsu,Sauvola,and other methods.