Crop Insect Recognition And Counting Based On Convolutional Neural Network

Crops are vulnerable to various plant diseases and insect pests during growth,which affects the yield and quality of grain.Therefore,it is particularly important to control crops' pests and diseases.The premise of effective pest control is timely and accurate pest monitoring in the field.Monitoring the population dynamics of pests based on the insect image is one of the effective control methods.Because the ecological environment of agricultural pests is complex,the species and quantity are huge,the traditional monitoring methods based insect image method has the problems of poor timeliness and low accuracy.In order to solve this problem,this paper uses deep learning image processing technology and proposes a set of image segmentation and image recognition methods based on convolutional neural networks.Completed automatic recognition and counting of crop insect images which collected in the farmland.The main research of this paper include:(1)Research on background segmentation.Due to the complexity of the image background,the Fully Convolutional Network was used to separate the insects from the background.Based on the U-Net model,this paper constructs a model named Insect-Net for Insect image segmentation.The model extracts features from the complete insect picture and the cut insect picture,and fuses them.The fused features will pass through a 1 × 1convolution layer to obtain the final segmentation result,and then extract the insect contour and count them.The experimental results show that this method achieves higher pixel accuracy and count accuracy in insect image,which are 94.4% and 89.2%.(2)Research on adhesion segmentation methods.The contours of insects separated from the background are partially adhesion.In order to accurately recognize and count,the adhesion insects must be separated.In this paper,we used Insect-Net model to train a sub-model using the insect target area as label and another sub-model using the insect target edge as label.Finally,do AND operation on the binary results predicted by two model.We called this model the two-stream Insect-Net model.Compared with traditional segmentation methods,two-stream Insect-Net has better segmentation rate and segmentation efficiency rate.(3)Research on insect classification methods.The Insects are divided into large insects and small insects in the single-target model,different feature extraction networks are used for each of them to compare the classification effects.The results show that VGG4 model performed better on small insects and Res Net50 model better on large insects.However,theCRNN model combining the adhesion segmentation task and insect recognition task is not advantageous in the case that a large number of samples are single target.(4)We combined the above three methods to compare the experiment results of crop insect recognition and counting.The results showed that the combination method with CRNN model had best counting accuracy,and its average counting accuracy of all kinds of insects was 86.9%.In this paper,an automatic identification and counting model is established for six kind of target pests.The experimental results show that the model has better accuracy and generalization ability,and can satisfy the needs of pest detection in agricultural production environment.