摘要
花卉识别在生活中有重要的应用和研究价值。目前传统的花卉识别方法都是通过人工手动选择多个特征进行分类,存在识别准确率低、泛化能力较弱和分类时间长等问题。由于不同的花朵之间存在相似度,通过对每张图片随机变化,增强数据集,把花卉作为研究对象,提出了一种基于ResNet18网络模型优化的花卉识别方法。将ResNet18网络模型中残差块的基础卷积替换为空洞卷积,提取花卉图片更多的细节特征来实现高精度,接着在每个残差块后分别加入经过改进的通道注意力机制优化网络权重,构造改进的ResNet18网络模型,在Oxford 102 Flowers牛津花卉数据集上的实验进行了仿真。实验结果显示,在Oxford 102 Flowers牛津花卉数据集上ResNet网络模型相较于基础AlexNet、VGG-16网络模型准确率高。改进的ResNet网络模型识别精度可以高达97.78%,比仅使用空洞卷积的模型提高了3.11个百分点,比原模型提高了4.45个百分点。改进的ResNet18网络模型在花卉识别的泛化和拟合能力有显著的提高。
Flower recognition has important application and research value in life.At present,the traditional flower recognition methods select multiple features manually for classification,which has some problems,such as low recognition accuracy,weak generalization ability and long classification time.Due to the similarity between different flowers,by randomly changing each picture and enhancing the data set,taking flowers as the research object,a flower recognition method based on ResNet18 network model optimization is proposed.The basic convolution of residual blocks in ResNet18 network model is replaced by void convolution,and more detailed features of flower pictures are extracted to achieve high precision.Then,after each residual block,an improved channel attention mechanism is added to optimize the network weight,and an improved ResNet18 network model is constructed.The experiment is simulated on Oxford 102 Flowers data set.The experimental results showthat the accuracy of ResNet network model is higher than that of basic AlexNet and VGG-16 network models on Oxford 102 Flowers data set.The recognition accuracy of the improved ResNet network model can be as high as 97.78%,which is 3.11 percentage points higher than the model using only hole convolution and 4.45 percentage points higher than the original model.The generalization and fitting ability of the improved ResNet18 network model in flower recognition are significantly improved.
作者
赵洋
梁迎春
许军
李大舟
ZHAO Yang;LIANG Ying-chun;XU Jun;LI Da-zhou(School of Computer Science and Technology,Shenyang University of Chemical Technology,Shenyang 110142,China;Key Laboratory of Industrial Intelligence Technology on Chemical Process of Liaoning Province,Shenyang 110142,China)
出处
《计算机技术与发展》
2022年第7期167-172,共6页
Computer Technology and Development
基金
辽宁省教育科学研究项目(L2016011)。
