摘要
针对传统图像分类方法在花卉图像上存在分类效果不佳的问题,提出一种改进Xception网络的方法。首先结合Res2net中的多尺度模块来提高模型特征信息的丰富度,提出Multi_Xception网络,接着使用1×1卷积核对多尺度深度可分离卷积模块的输入特征图进行信道压缩,减少模型参数的同时进一步丰富模型特征信息,提出Multi2_Xception网络。将改进模型应用于Flowers Recognition花卉数据集分类,实验结果表明,该方法相较于原算法分类准确率提升了1.64%,F1-score提升了0.018,验证了多尺度Xception网络的有效性。
Aiming at the problems of poor classification effect traditional image classification methods in flower images,applied the Xception network with obvious classification effect and fast convergence to flower image classification and an improved xception network method is proposed.This method first uses the Xception network with an obvious classification effect and fast convergence as the basic network.Then,proposes the Multi_Xception network by combining the multi-scale module in Res2net to improve the richness of model feature information,and finally the 1×1 convolution kernel are used to compress the input feature map of the multi-scale depthwise separable convolution module to reduce the model parameters and further enrich the model feature information,and Multi2_Xception network is proposed.The improved model is applied to the classification of flowers recognition data set,through experiments,the classification performance of this method has increased 1.64%and F1 score has increased 0.018 than original classification methods.Verification,the effectiveness of the multi-scale Xception network.
作者
赵正伟
朱宏进
ZHAO Zhengwei;ZHU Hongjin(College of Electronic Information,Guangxi Minzu University,Nanning 530006,China)
出处
《广西民族大学学报（自然科学版）》
CAS
2023年第2期90-96,共7页
Journal of Guangxi Minzu University ：Natural Science Edition
基金
广西混杂计算与集成电路设计分析重点实验室开放课题一般项目(HCIC201511)
广西民族大学高等教育改革项目(2020XJGY41)。
