Research On SSD-based Method Applied To Grape Leaf Disease Identification

The disease of grape leaf has negative effect on its yield and quality,it is thus very important that grape leaf disease is identified timely and controlled in the cultivation.With the continuous developing of artificial intelligence technology,convolutional neural networks are applied to the automatic identification of the disease of grape leaf by many researchers.The precise identification of convolutional neural networks can be carried out by accurate feature extraction of the leaf,including color,texture,shape,and so on.However,each of the convolutional neural networks perform differently in classification accuracy,computer resource,operation efficiency,and so on.Based on the traditional convolutional neural network and single shot detector network(SSD),a novel object detection network named as EGR＿SSD is proposed,which obviously enhances the accuracy of disease classification and marks the disease position.Firstly,the VGG16 of the single shot detector network(SSD)is substituted by the ResNet50 and the original four cascaded convolution layers are substituted by some residual blocks.The above substitutions deepen the network and enhance disease feature extraction.Secondly,by reducing the number of the convolution kernel of stage＿5 in the ResNet50,the channel number is decreased from 2048 to 512.It is helpful to model compression and save computer resource.The original normalization module of Batch Norm is substituted by Group Norm,which overcome the limit of batch size,specially,of small batch size.At last,efficient channel attention(ECA)is introduced in EGR＿SSD to enhance efficiency of feature extraction.In order to evaluate the disease identification performance of EGR＿SSD,a dataset named as GLDDS,which includes 2100 images of disease leaves,is fabricated from open Plant Village dataset by Label Img tool.Using the GLDDS,some comparison experiments are carried out among proposed EGR＿SSD,SSD,Faster R-CNN,YOLOv3,and SSD+ResNet50.The results show that EGR＿SSD has the best disease identification performance from the aspects of value of AP and m AP.For black rot,black measles,and leaf blight,the AP obtained by EGR＿SSD are respectively 90.11%,64.20%,76.23%.Further,the obtained m AP is about 76.85%,which is higher12.15%,6.05%,9.31%,7.64%than Faster R-CNN,YOLOv3,and SSD + ResNet50 respectively.