基于深度学习的自然场景文本检测方法的特征提取模块一般采用大型网络,模型复杂且效率低.为了降低文本检测模型的复杂度以及更快速有效地检测文本,在基于分割的渐进式扩展网络PSENet的基础上,使用轻量级小型网络MobileNet V3作模型局部特征提取模块,减少参数数量,结合多级卷积来提取不规则文本的区域特征;使用优化器Adam计算每个参数的自适应学习率,加速训练优化过程,提升模型运算效率.在数据集ICDAR2015上进行验证,实验结果表明改进的算法在性能上有明显改善.
The characteristic extraction module based on deep learning is generally large network,which is generally large net-work with complex and low efficiency.In order to reduce the complexity of the text detection model and detect the text faster and effec-tively,on the basis of a divided gradient expansion network PSENet,use lightweight small network MobileNet V3 as a local feature ex-traction module to reduce the number of parameters,combined with multiple combination class convolution to extract the regional char-acteristics of irregular texts;use the optimizer Adam to calculate the adaptive learning rate of each parameter,accelerate the training optimization process,and improve the model operation efficiency.Verification on the data set ICDAR2015,the experimental results show that the improved algorithm has improved significantly in terms of performance.