Accurate segmentation of blood vessels in retinal images facilitates the observation of ocular lesions.In order to im-prove the retinal image blood vessel segmentation accuracy and feature information reuse rate and simplify the model,starting from the network framework,a combination of DCSAU-Net(Deeper and Compact Split-Attention U-Net),multi-scale information fusion block and Ghost module is proposed.Retinal image vessel segmentation model-MLDCSAU-Net(Multi-scale Lightweight Deeper and Compact Split-Attention U-Net)model.The model improvement mainly includes two aspects:first,the multi-scale information fusion block is introduced after the skip connection;second,the encoder uses the Ghost module to replace the encoder end CSA block.The experimental results show that:the multi-scale information fusion block has greatly improved the segmentation accuracy of the model;the Ghost module has effectively reduced the number of model parameters.In the three public data sets of STARE,CHASEDB1 and HRF,the accuracy,precision,recall and F1 score of the MLDCSAU-Net model are higher than those of the origi-nal model,and the number of parameters is smaller.
关键词
视网膜图像血管分割/多尺度信息融合模块/Ghost模块
Key words
retinal vascular segmentation/multi-scale information fusion block/Ghost module
维普
万方数据