卷积神经网络下的相似月季识别  

Recognition of similar rose based on convolution neural network

作　　者：王雪琰 张冲 张立[1] WANG Xueyan;ZHANG Chong;ZHANG Li(College of Science,Beijing Forestry University,Beijing 100083)

机构地区：[1]北京林业大学理学院,北京100083

出　　处：《安徽农业大学学报》2021年第3期504-510,共7页Journal of Anhui Agricultural University

基　　金：中央高校基本科研业务费专项基金(2019SG04)资助。

摘　　要：目前我国关于花卉分类识别技术已较为成熟,但由于类间相似性较高,花卉特征较难提取,对于同种不同类的相似花卉仍存在识别率较低的问题,因此提出使用卷积神经网络(CNN)中4类深度学习网络SqueezeNet、ResNet、InceptionV3及DenseNet的训练模型,搭建了对4种相似月季进行识别的花朵识别客户端,并对识别结果进行比较,筛选出最优模型,同时运用GPU对训练过程以及识别过程进行加速。对实验过程产生的数据进行统计对比后得出Inception V3网络训练后得到的模型较其余3种网络而言识别率最高且识别速度较快,可以作为最优模型。将搭建的花朵识别系统应用于花卉分类工作中,在节省人工的同时也能够加速园艺自动化的进程。At present,the classification and recognition technology of flowers in China is relatively mature.However,due to the high similarity among species,the flower characteristics are difficult to extract.This research proposed to use the training models of four kinds of deep learning networks,SqueezeNet,ResNet,InceptionV3 and DenseNet in Convolution Neural Network(CNN)to build flower recognition clients for recognizing four kinds of similar roses,to compare the recognition results,so as to screen out the optimal model and to accelerate the training process and recognition process by GPU.After statistical comparison of the data generated in the experimental process,it was concluded that the recognition rate and recognition speed of the model trained by Inception V3 network are the highest and the recognition speed is faster than the other three networks,which can be used as the optimal model.The application of the flower recognition system in flower classification can not only save labor,but also accelerate the process of gardening automation.

关 键 词：月季 SqueezeNet ResNet InceptionV3 DenseNet 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]