基于双卷积链Fast R-CNN的番茄关键器官识别方法

Recognition Method of Tomato Key Organs Based on Dual Convolution Fast R-CNN

为提高番茄器官目标识别的准确率,提出一种基于RGB和灰度图像输入的双卷积链Fast R-CNN番茄器官识别网络。该方法通过番茄器官图像数据集训练基于VGGNet基本结构的特征提取网络,并用其参数初始化Fast R-CNN,通过再训练,用以识别植株图像中的番茄花、果、茎器官。首先分析了网络深度和宽度、图像输入类型、激活单元对特征提取及网络分类性能的影响。详细阐述了基于Fast R-CNN的番茄器官识别网络的设计及训练方法,基于试验观察,提出了基于双卷积链的Fast R-CNN,融合自动提取的RGB和灰度图像特征,由全连接层对Selective Search算法生成的候选区域进行分类识别。结果表明:针对番茄器官图像数据集,5个卷积层的网络即可具有较高的特征提取和分类性能,增加或降低卷积层数都会使网络性能下降;与ReLU激活单元相比,PReLU和ELU能够显著提高番茄特征提取网络的性能,而提高效果和具体的网络结构有关;基于Fast R-CNN的识别方法能够对番茄的花、果、茎器官进行识别,且能够识别不同成熟度的果和不同形态的花;单卷积链Fast R-CNN网络对花、果、茎的识别平均精度(AP)最高分别为64.79%、66.76%和42.58%,双卷积链Fast R-CNN识别网络对三种器官的识别AP最高分别为70.33%、63.99%和44.95%,相较于单链网络,双卷积链Fast R-CNN的mAP提高2.56%,说明该方法对提高番茄器官识别性能是有效的。

In order to enhance the proportion of directional features in classification features and improve the recognition accuracy of main organs of tomato,a dual convolutional Fast R-CNN tomato organ recognition network based on RGB and grayscale image inputs was proposed.This method used the tomato organ image data set to train the feature extraction network based on the basic structure of VGGNet,and used its parameters to initialize Fast R-CNN,and then retrained Fast R-CNN to identify tomato flowers,fruits,and stem organs in plant images.The effects of network depth and width,input image type and activation unit on feature extraction and classification performance were analyzed.The design and training method of tomato organ recognition network based on Fast R-CNN were described in details.Based on the experimental observation,a dual convolutional Fast R-CNN network was proposed to fuse the extracted RGB and grayscale image features.The object region proposals which were generated by the Selective Search algorithm were classified and recognized by the three Fully-Connected layers.The experimental results indicated that the 5 convolutional layers network could have high feature extraction and classification performances for tomato organ image datasets,and that increasing or decreasing the number of convolution layers would reduce the network performance.Compared with the ReLU activation unit,the PReLU and ELU unit could significantly improve the performance of tomato organ feature extraction network,and the improvement effect was related to the specific network structure.The identification method based on Fast R-CNN could identify the flowers,fruit and stem organs,as well as the fruits of different maturities and different forms of flowers.The average accuracies(APs)of single convolution chain Fast R-CNN network for flower,fruit and stem recognition were 64.79%,66.76%and 42.58%respectively,and the APs of dual convolution Fast R-CNN for three types of organs were 70.33%,63.99%and 44.95%,respectively.Compared with single convolutional network,the mAP of dual convolutional Fast R-CNN increased by 2.56%.The result showed that this method is effective to improve the tomato organs recognition performance.