首页|基于窗口注意力机制的文本超分辨率方法

基于窗口注意力机制的文本超分辨率方法

扫码查看
自然场景下文本图像往往具有复杂且多样化的背景,由于拍摄条件的限制,这些图像可能存在模糊、昏暗或低分辨率等问题,给文本识别任务带来了挑战.为了提高文字区域的清晰度并提升文字识别的准确性,提出了文本超分辨率网络TSSRN.该算法在文本超分辨率网络TBSRN的基础上引入了Swin Transformer,以感知更多层次的特征信息.此外,还引入了边缘损失函数增强纹理结构特征.最后在TextZoom数据集上进行训练和测试,实验结果表明,该方法在文本图像超分辨率重建的细节清晰度方面取得了一定的提升,相较于原TBSRN算法也有所提高,从而验证了算法的有效性.
Swin Transformer Based Text Super-Resolution Method
Text images in natural scenes often have complex and diverse backgrounds.Due to the limitations of shoot-ing conditions,these images may have issues such as blurring,dimming,or low resolution,which poses challenges to text recognition tasks.In order to improve the clarity of text regions and improve the accuracy of text recognition,this paper proposes a text super-resolution network TSSRN.This algorithm introduces Swin Transformer on the basis of text super-resolution network TBSRN to perceive more levels of feature information.In addition,the edge Loss function is introduced to enhance texture structure features.Finally,this paper conducted training and testing on the TextZoom dataset,and the ex-perimental results showed that the method has achieved certain improvements in super-resolution reconstruction of text im-ages,which is also improved compared to the original TBSRN algorithm,thus verifying the effectiveness of the algorithm.

super-resolution reconstructiontext recognitiondeep Learning

舒琪、王旭智、万旺根

展开 >

上海大学通信与信息工程学院,上海 200072

上海大学智慧城市研究院,上海 200072

超分辨率重建 文本识别 深度学习

2024

工业控制计算机
中国计算机学会工业控制计算机专业委员会 江苏省计算技术研究所有限责任公司

工业控制计算机

影响因子:0.258
ISSN:1001-182X
年,卷(期):2024.37(4)
  • 7