摘要
为提升用于眼科超声图像数据集的卷积神经网络分类效果,对于目前眼科超声数据集存在图像低对比度干扰、有错分漏分等问题,在GoogleNet分类网络模型的基础上,设计了一种改进的眼科超声图像分类方法,该方法引入递归门控卷积以及由不同大小卷积组组成的多尺度结构用来增强网络模型特征提取能力.在眼科超声数据集上进行了大量实验.实验结果表明,与原GoogleNet网络模型对比准确率、宏精确率、宏召回率和宏F1分数,分别提高了1.41%、4.36%、5.02%、5.14%.与EfficientNetB0、ResNet50等主流分类网络相比,准确率分别提升了2.21%、5.03%.
Abstract
In order to improve the classification effect of convolutional neural network used in ophthalmic ultrasound image datas-ets,there are problems such as low image contrast interference,wrong classification and omission of classification in the current oph-thalmic ultrasound datasets.Based on the GoogleNet classification network model,an improved ophthalmic ultrasound image classifica-tion method is designed,which introduces recursive gated convolution and a multi-scale structure composed of convolution groups of different sizes to enhance the feature extraction ability of the network model.Extensive experiments are performed on ophthalmic ultra-sound datasets.The experimental results show that compared with the original GoogleNet network model,the accuracy rate,macro pre-cision rate,macro recall rate and macro F1 score have increased by 1.41%,4.36%,5.02%,and 5.14%,respectively.Compared with mainstream classification networks such as EfficientNetB0 and ResNet50,the accuracy rates have increased by 2.21%and 5.03%,re-spectively.
维普
万方数据