改进PSENet的自然场景文本检测方法

Improved PSENet natural scene text detection method

基于深度学习的检测方法在文本形状较规则的情况下,已经取得较好的检测结果,但对于倾斜以及弯曲的文本行仍有改进空间。文章在渐进式尺度扩展网络PSENet的基础上,通过使用Res2Net模块提取多尺度特征,并结合全局卷积网络GCN进行特征融合,来对原有模型进行改进。根据在SCUT-CTW1500和Total-Text数据集的实验结果对比,证明改进的算法有效可行。

The detection method based on deep learning has achieved good results in the case of regular text shape,but there is still an improved space for tilting and curved text.Based on the progressive scale expansion network PSENet,the original model is improved by using the Res2Net module to extract multi-scale features and combined with the global convolution network GCN.According to the comparison of experimental results in SCUT-CTW1500 and Total-Text data sets,the effectiveness of the improved algorithm is proved.