摘要
植物分类在形态、颜色和纹理上具有高度的相似性和密集的细节信息,传统的机器学习方法无法满足这些大样本的特征提取训练,识别种类与精度受到限制。深度学习可以有效地解决植物图像识别在种类数量、准确度和速度上的难点。本文提出了基于优化的P-AlexNet模型的植物识别算法,基于卷积神经网络(CNN)中的AlexNet网络模型进行优化处理,提高模型的泛化能力、细节特征的表征能力以及识别精度。利用迁移学习热启动更新植物识别种类,利用GPU并行计算加速模型训练和图片识别速度。针对206类植物图片,训练得到验证集精度达到86.7%的模型。以此模型为基础,开发了一款智能植物图像识别平台,包含了Web网站以及Android和IOS的App应用。Web端实验测试结果表明,检测时间平均为1.282s,具有较高的准确性和泛化性以及快速的识别速度。
Plants have high similarity and detail information in morphology,color,and texture.Traditional machine learning methods cannot meet the demands of feature extraction from many plants and they only recognize several types of plants,while deep learning can effectively deal with these difficulties,including amount,accuracy and speed of plant images recognition.This paper proposes a plant recognition algorithm based on the optimized P-AlexNet model.The traditional AlexNet mainly considered the classification of images with large difference between targets and ignored the difference of plant images.Therefore,more attention should be paid to distinguish deeper features when designing the network structure.By using inception module instead of traditional convolutional pooling single-channel structure,the proposed model can increase the representable range of underlying texture features.The green channel is separated from the remaining two channels such that the extracted features can characterize the information of leaf texture and the structure of the flower to attain better interpretability of the network.The contrastiveloss function in the siamese network is utilized to improve the recognition accuracy among plant categories after full connection layer.Based on the AlexNet network model in CNN,the generalization and the characterization of the detail features and the recognition precision of the model can be effectively improved.The concept of migration learning is employed to update the plant identification and GPU parallel computing is adopted to speed up model training and image recognition speed.The model training uses an image dataset with206plants,composed of Oxford102and Ecust104dataset,and the validation accuracy of the model is86.7%.Based on this proposed model,this paper further develops a platform for intelligent plant image recognition,including Web site and App application of Android and IOS.It is shown from these experiments that the average detection time is1.282s and the higher accuracy and generalization and fast recognition speed can be attained.
作者
张雪芹
陈嘉豪
诸葛晶晶
余丽君
ZHANG Xue-qin;CHEN Jia-hao;ZHUGE Jing-jing;YU Li-jun(School of Information Science and Engineering,East China University of Science and Technology,Shanghai 200237,China)
出处
《华东理工大学学报（自然科学版）》
CAS
CSCD
北大核心
2018年第6期887-895,共9页
Journal of East China University of Science and Technology
基金
国家自然科学基金(31671006)
