Insect Recognition Method Based On Deep Learning And Its Application

In recent years,with the development of agriculture in the direction of digitization and precision,traditional insect identification methods are difficult to meet the development needs of modern agriculture.Deep learning has obvious advantages in image feature extraction and modeling,and has achieved good results.Therefore,this paper adopts the method of deep learning,based on the Faster R-CNN target detection algorithm,and proposes an automatic insect recognition method based on deep learning.In this paper,the insect image acquisition and processing,the feature extraction network model selection,the Faster R-CNN model optimization and the improvement sequence are carried out,and the theory and experiment are combined to compare and analyze.The main work and research contents are as follows:(1)Firstly,the research background,significance and research status of crop insect identification technology are introduced.The deep learning theory is elaborated.The layers and working principles of convolutional neural networks are introduced.The development,detection process and related calculations of the convolutional neural network Faster R-CNN are described in detail.Three insect images of Datun,Chilosuppressalis and Diamond were collected and labeled,and a total of 1643 original sample data were collected.The original dataset was performed by image rotation,mirror transformation,image saturation,brightness and image Gaussian addition.Enhanced,the sample size was expanded to 2390 sheets,and an insect VOC data set was created according to the data format requirements of Faster R-CNN;(2)The basic network models of VGG16 and ResNet101 are introduced in detail.Through theoretical analysis and experimental comparison,it is found that Faster R-CNN based on ResNet101 has better recognition effect than VGG16 on insect dataset,and the accuracy rate is improved by 0.03 mAP.The average recognition accuracy of the three pests reached 93%,indicating that the detection algorithm using the ResNet101 model can effectively identify and extract the insect targets in the image.Therefore,ResNet101 was selected as the front-end network for the subsequent insect identification research;(3)Under the framework of Faster R-CNN algorithm,three common network optimization algorithms,such as stochastic gradient descent,momentum and adaptive momentum optimization,are analyzed and compared.The results show that the adaptive momentum optimization algorithm can adjust the learning according to the change of parameters.Rate,speed up the convergence of the model;(4)The method of dynamically adjusting the weight gradient based on loss transform is proposed to improve the Momentum algorithm,and the method is verified by experiments.To some extent,the network oscillation can be effectively reduced,and the network can be prevented from falling into local optimization.Improvements to Faster R-CNN.