基于RDEB-FCE模型的西夏文本检测研究OA

A Xixia text detection method RDEB-FCE based on dilated convolution and attention guidance is proposed to address the problems of missed and false detections caused by the diverse scales and irregular shapes of text instances in current Xixia text detection tasks.The mod-el first uses Resnet50 as the backbone network and employs FPN structure to capture multi-scale features;Secondly,utilizing dilated convolu-tion to expand the receptive field of features and enhance the multi-scale capture capability of feature information,while adopting an efficient channel attention mechanism to adaptively adjust the weights of channel features,in order to improve the performance of the model in process-ing large-scale and high-resolution Western Xia text images;Finally,in the regression loss function,Smooth-L1 loss is replaced with Bal-anced L1 loss to improve the gradient of accurate samples and thus enhance the accuracy of text detection.The experimental results show that the accuracy of this method on the Xixiawen dataset constructed in the laboratory reaches 92.4％,which is a significant improvement compared to the current mainstream methods.