摘要
使用卷积神经网络进行图像识别,可以大大降低图像辨识的成本,在二分类问题中尤其如此。VGG模型是一种相当流行的卷积神经网络,其特性在于以小卷积核和“网络块”替代传统神经网络中的大卷积核与神经网络层,这意味着其深度有所增加,同时具有较强的迁移性与改进潜力。通过测试发现,增加VGG块数的同时,搭配图像增强是可靠的改进手段;而增加epoch有利有弊,对网络进行dropout的成效不太理想。针对于此,实验构建了一个准确度为83.3%的轻量化VGG模型,该模型相较VGG-16而言要轻量化许多,表明根据合理的方向构建轻量化VGG模型用于动物识别是可行的。
Using CNN for image recognition can greatly reduce the cost of image recognition,especially in classification problems.VGG model is a very popular convolution neural network.Its characteristic is that small convolution kernel and"network block"are used to replace the large convolution kernel and neural network layer in the traditional neural network,which means that its depth is increased and has strong improvement potential.Through the test,we found that increasing the number of VGG blocks and image enhancement are reliable means of improvement.Adding epoch has advantages and disadvantages,but the effect of dropout on the network is not ideal.Finally,this project has built a lightweight VGG model with an accuracy of 83.3%,which is much simpler than VGG-16,which shows that it is feasible to build a lightweight VGG model for animal recognition.
作者
龚建伟
孟博文
童昱恒
孔煜杰
谭仪慧
GONG Jianwei;MENG Bowen;TONG Yuheng;KONG Yujie;TAN Yihui(Faculty of Science and Technology,Beijing Normal University-Hong Kong Baptist University United International College,Zhuhai Guangdong 519087,China)
出处
《智能计算机与应用》
2023年第5期70-74,81,共6页
Intelligent Computer and Applications
关键词
动物识别
VGG
卷积神经网络
优化路径
animal recognition
VGG
convolutional neural network
optimization