【摘要】 针对现有场景文本图像超分辨率方法存在真实世界高分辨率(high-resolution, HR)-低分辨率(low-resolution, LR)配对训练图像收集困难，文本语义先验信息利用不充分的问题，提出一种改进的文本语义先验引导的场景文本图像超分辨率重建方法(enhanced text prior guided scene text image super-resolution, ETPGSR)。将文本先验生成器从卷积循环神经网络(convolutional recurrent neural networks, CRNN)改进为DenseNet-RNN;改进语义先验特征变换模块，通过5层转置卷积将特征图尺寸扩大到接近超分辨率分支中特征图尺寸；引入文本先验特征与图像超分辨率特征融合模块；提出可学习神经算子与常规算子结合的场景图像降质模拟方法以构建HR-LR配对的数据集。实验结果表明：通过模拟数据集的预训练结合改进的文本先验分支网络结构，使用ASTER(attentional scene text recognizer with flexible rectification)、MORAN(multi-object rectified attention network)和CRNN对超分辨率重建后图像进行文本识别，在TextZoom数据集上分别取得了64.5%、60.8%和54.0%的准确率，超过TPGSR(text prior guided super-resolution)、TATT(text attention network)等多个对照模型；在ICDAR2015和SVT数据集上的泛化性测试结果同样超过上述对照模型。可见ETPGSR能有效提升文本图像超分辨率的性能。更多还原

【Abstract】 To address the challenges of collecting real-world high-resolution(HR) and low-resolution(LR) paired training images and the insufficient utilization of text prior information in existing scene text image super-resolution(STISR) methods, an improved method named enhanced text prior guided scene text image super-resolution(ETPGSR) was proposed. The text prior generator was upgraded from convolutional recurrent neural networks(CRNN) to DenseNet-RNN. The text prior transformer was improved by expanding the feature map size through a 5-layer transposed convolution to match the dimensions of the feature maps in the super-resolution branch. A fusion module was introduced to combine text prior features with image super-resolution features. Additionally, a scene image degradation simulation method combining learnable neural operators and conventional operators was proposed to construct a HR-LR paired dataset. Experimental results show that with pretraining on simulated dataset and an improved text prior branch network architecture, text recognition accuracy of 64.5%, 60.8%, and 54.0% is achieved using attentional scene text recognizer with flexible rectification(ASTER), multi-object rectified attention network(MORAN), and CRNN, respectively, on the benchmark TextZoom dataset. These results surpassed multiple comparison models such as text prior guided super-resolution(TPGSR) and text attention network(TATT). Furthermore, generalization tests conducted on the ICDAR2015 and SVT datasets also demonstrate superior performance over the aforementioned models. It is concluded that ETPGSR can effectively enhance STISR performance.更多还原