Application Of Target Detection Algorithm Based On Deep Learning In Detection And Recognition Of Farmland Pests

In order to carry out real-time and efficient monitoring of pests and diseases occurring in farmland,there is a key technology that needs to be broken through,which is to realize automatic detection and recognition of farmland pest images.There are many methods to detect farmland pests.Based on the deep learning target detection technology,this thesis proposes an improved convolutional neural network with powerful feature learning ability to implement farmland pest image recognition.In this thesis,the common pests in farmland are taken as the research object.By combining theoretical analysis and experimental verification,the research work is carried out in the order of farmland pest data set production,selection and design of detection model,and optimization of detection model.The thesis carries out the research in the way of theoretical analysis and experimental comparison.(1)A total of farmland pest images containing 10 specieses are collected,and a total2472 samples are annotated.The collected sample set is expanded by using the basic operations such as rotation and translation of the original image and adding noise.Then sample capacity is expanded to 12474.The sample set has certain generality and can be used in the technical research for farmland pest identification and image recognition.(2)Aiming at the research characteristics of farmland pest recognition,Faster R-CNN is selected as the target detection algorithm in this thesis.In order to obtain better detection accuracy and speed,the basic feature extraction network of Faster R-CNN is changed to a more efficient DenseNet network to design a farmland pest detection model.In the follow-up farmland pest detection experiment,the detection effect of the model is verified,and the accuracy reaches 90.78%.(3)After analyzing and summarizing the defects of the Faster R-CNN(DenseNet)model,the optimization of the model is carried out.In terms of data sets,the original data sets are expanded again,and the sample capacity reaches 24948.In terms of target detection algorithm,a more optimized RefineDet algorithm is selected.At the same time,a new ARM network architecture which combined with newly constructing anchor frame is proposed to design the model of farmland pest detection.Experimental results show that two optimization methodscan improve the performance of insect recognition,the model achieves a recognition rate of92.78%.(4)This thesis optimizes the random gradient descent algorithm used in the research,and forms a training algorithm which can automatically reduce the learning rate.This algorithm can detect whether the training of the network has reached saturation at the current learning rate,and if so,it will automatically reduce the learning rate and continue to complete the training.And the optimized algorithm is verified through experiments,which improves the efficiency of training,the learning rate does not have to be adjusted through the step size during training,and the final recognition rate is also improved.Finally,the adaptive learning rate descent algorithm is applied to the training of farmland pest detection models.The final optimized RefineDet(DenseNet-121)farmland pest detection model reaches a recognition rate of 93.83% in the experiment.