视网膜血管分割全卷积网络的数据集预处理研究
首发时间:2019-03-28
摘要:在深度神经网络中,对于数据集的预处理工作是非常重要的,会直接影响神经网络模型的预测结果。对于视网膜血管分割任务而言,数据集的预处理工作是否、以及怎样对卷积神经网络的分割结果造成影响,是人们需要认真考虑的问题。本文在使用一个U-Net全卷积网络对DRIVE数据集的彩色眼底图像进行视网膜血管分割任务时,尝试了许多不同的数据集预处理策略,并通过对比实验比较了这些预处理方法对最终结果的影响。本文确定了使用0.5几率随机翻转、在原始图像上随机采样图像块、不对归一化输入数据进行标准化会取得最好的分割效果。
关键词: 模式识别与智能系统 卷积神经网络 语义分割 视网膜血管分割
For information in English, please click here
The Preprocessing of Retinal Vessel Segementation Dataset for Fully Convoulutional Network
Abstract:The preprocessing of datasets is important for deep neural networks, which can change the performance of network significantly. It should be treated seriously by researchers as the preprocessing of datasets may be influence the segmentation result in the tesk of retinal vessel segmentation tesk. In this paper we use a fully convolutional network with U-Net architecture for the vessel segmentation tesk of DRIVE retina datasets. We have tested some different data preprocessing strategy, and compared the results of these strategies and their influences. The proposed preprocessing strategy is that crop the image patch randomly and flip it with 0.5 probability, and do not apply normalization of uniformizationed input data, which can achieve better performance.
Keywords: Pattern Recognition and Intelligent System Convolutional Neural Network Semantic Segementation Retinal Vessel Segmentation
基金:
引用
No.****
动态公开评议
共计0人参与
勘误表
视网膜血管分割全卷积网络的数据集预处理研究
评论
全部评论0/1000