Research On Image Identification Of Crop Disease And Weed Based On CNN And Transfer Learning

Automatic identification of crop diseases and weed types is of great significance for improving crop yield and quality.Digital image processing for crop disease and weed identification usually involves image segmentation.When the field light changes or the background is complex,it will affect the image segmentation effect,and thus reduce the subsequent identification rate.In addition,this method requires a large number of feature extraction,and high computational complexity will reduce the identification efficiency.Aiming at these problems,this dissertation applies convolutional neural network algorithm for automatic identification of crop disease and weed.Based on the convolutional neural network algorithm,this dissertation introduces the transfer learning strategy to alleviate the over-fitting phenomenon for the problem that the massive samples in the agricultural field are difficult to obtain and label.The research work and innovations of this dissertation are as follows:(1)For the over-fitting problem generated by training deep network for small data samples,the crop disease and weed identification methods based on shallow convolutional neural network are studied.On the one hand,the impact of network depths on classification results is discussed.By constructing CNNs with different depths,we studied the network model suitable for our disease and weed dataset,and completed the automatic extraction of crop disease and weed characteristics.On the other hand,the influence of training mechanism on classification results is explored.For the over-fitting problem caused by small dataset,we used the related PlantVillage dataset to obtain the pre-trained model,and adjusted the parameters to adapt to our dataset.(2)The single CNN identification model has its own strengths and can be combined to improve the identification accuracy.This dissertation studies crop diseases and weed identification methods based on deep CNN model integration and parameter fine-tuning.Firstly,data enhancement technology is used to expand the dataset size.Then,this dissertation makes full use of four single deep CNN network frameworks to train and fine-tune the parameters based on the prior knowledge learned from the big dataset,so as to alleviate the over-fitting problem caused by insufficient data sources.Finally,the accuracy of crop disease and weed identification is further improved by training multiple neural network models and synthesizing their predictions with the direct average method and the weighting method.(3)Aiming at the problem that the identification speed of deep CNN models and parameter fine-tuning strategy is slow and not conducive to practical application,the crop disease and weed identification system based on CNN bottleneck layer feature extraction is proposed.The system combines deep CNN models with transfer learning strategy based on feature extraction of the bottleneck layer.Inception-v3 and Mobilenet with different parameters are used as the underlying feature extractor for feature extraction,and then crop diseases and weeds are identified through the classification model.Based on this theory,we developed the crop disease identification system based on Mobilenet model and the corn weed identification system based on Inception-v3 model.