基于Faster R-CNN的苦瓜叶部病害识别研究

Research on Recognition of Bitter Melon Leaf Diseases Based on Faster R-CNN

为实现对苦瓜叶部多种病害的自动检测,提出了基于Faster R-CNN的目标检测算法对自然环境条件下的苦瓜健康叶片和叶部病斑进行目标检测。该算法利用了预训练的ImageNet深度网络模型进行迁移学习,以VGG-16卷积神经网络作为本次试验的特征提取网络,并且结合苦瓜叶部病害尺寸小的特点,对原始的Faster R-CNN的参数进行修改,增加了训练时区域建议框的尺寸。结果表明,以VGG-16作为特征提取网络训练所得的深度学习网络模型对苦瓜健康叶片、白粉病、灰斑病、蔓枯病、灰斑病的平均精确率(Average Precision)分别为0.8849、0.8753、0.5539、0.8165、0.7520,平均均值精度(mean Average Precision,mAP)值为0.7765;增加候选框尺寸后,所得模型的mAP值为0.8014,提高了2.49%。该方法能够有效地实现对苦瓜叶部病害的分类与定位,对瓜果类疾病预防有重要的研究意义。

In this paper,the purpose of this paper is to realize the automatic detection of many kinds of diseases in balsam pear leaves,and a target detection algorithm based on Faster R-CNN is proposed to detect the diseases of balsam pear leaves in natural environment.In this algorithm,the pre-trained ImageNet depth network model is used for migration learning,and the convolution neural networks VGG-16 is used as the feature extraction networks of this experiment.Combined with the small size of bitter gourd leaf disease,the parameters of the original Faster R-CNN are modified to increase the recommended size of the area during training.The results showed that the deep learning network model trained with VGG-16 as feature extraction network had a better performance.The average accuracy rates of healthy leaves,powdery mildew,gray spot,vine blight and gray spot were 0.8849,0.8753,0.5539,0.8165,and 0.7520,the average mean precision(mean Average Precision,mAP was 0.7765.after increasing the size of candidate frame,the MAP value of the model was 0.8014,which was increased by 2.49%.This method can effectively realize the classification and location of balsam pear leaf diseases,and has important research significance for the prevention of melon and fruit diseases.