一种改进的深度神经网络的花卉图像分类

An Improved Depth Convolutional Neural Network for Flower Image Classification

花卉图像类内差异性大和类间相似性高使得花卉图像分类较难.传统花卉分类方法和普通卷积神经网络很难完整地表达花卉图像的特征,故而分类效果不理想.为提高花卉分类准确率,提出改进的InceptionV3网络用于花卉图片的分类.采用迁移学习的方法,将在大规模数据集上训练的InceptionV3网络用于花卉图像数据集的分类,对其中的激活函数进行改进.在通用Oxford flower-102数据集上的实验表明:该模型在花类图像分类任务中比传统方法和普通卷积神经网络分类准确率高,且比未改进的卷积神经网络准确率高,迁移过程准确率达到81.32%,微调过程准确率达到92.85%.

The difference within the same flower class and the high similarity among the different flower images make the classification of flower images more difficult.Traditional flower classification method and ordinary convolution neural network are difficult to express the characteristics of flower images,so the classification effect is not ideal.This article proposed an improved InceptionV3 network to classify flower pictures.The method of transfer learning was adopted.InceptionV3 network trained on large-scale datasets was used for the classification of flower image dataset,and the activation function was improved.Experiments on the Universal Oxford flower-102 dataset show that the proposed model has more advantages than traditional methods and general convolutional neural networks in the classification of flower images,and has a higher recognition rate than the unmodified convolutional neural network.The transfer-learning accuracy rate reaches 81.32%,and fine-tuning process accuracy rate reaches 92.85%.