Research On Identification Method For Rice Pests With Small Sample Size Problem Based On Deep Learning

The frequent outbreaks of rice pests affect the normal growth of rice as well as the quality and yield of rice.Identifying rice pests accurately is conducive to efficient insecticidal and control,which is the basis of health growth,high quality and high yield of rice.Currently,artificial identification is the main method to identify rice pests.This kind of judgment method has shortcomings such as strong subjectivity and low efficiency,which will cost a lot of manpower.With the development of computer vision technology in recent years,intelligent identification of rice pests has become possible.Limited by the shooting time and environment,the number of available samples of rice pest images in the real scene is small.On the basis of deep learning,this paper focuses on the identification of rice pests with small sample size problem in real scenes.The main work of this paper is as follows:1.A rice pest identification method combining deep and metric learnings is proposed in this paper.The U-Net network is used to segment the background of rice pest images and the largest connected region is screened out to optimize the background segmentation results.The VGG16 network is improved as a feature extraction network of rice pests,while the metric learning is introduced to identify the features of rice pests.The special U-shaped structure in the U-Net network can well retain the target information,and has a good segmentation effect on rice pest images with small sample size problem used in this experiment.The improved feature extraction network is suitable for feature extraction of rice pests with small sample size problem.Metric learning projects the feature vector of the pest image to new feature space for similarity matching and solves the problem that machine learning methods are difficult to accurately identify targets when the number of training samples is small.The results show that,in this experiment,which identifies rice pests with small sample size problem under the complex background,the method combining deep and metric learnings has a good identification accuracy.2.A rice pest identification method based on capsule network and Convolutional Block Attention Module is proposed.At the beginning,the Grab Cut is used to remove the complex environmental background in the rice pest images,while the Convolutional Block Attention Module(CBAM)is introduced into the capsule network to identify rice pests with small sample size problem.According to the threshold method,the smallest bounding box containing the pest is cropped,and part of the background is added to adjust the rectangular picture to a square.While removing the extra black background,it ensures that the pest shape will not be distorted when the image size is adjusted.The capsule network is mainly used to discover the relative position and angle of the features of rice pests more sensitively,and the CBAM can further improve the accuracy of the network’s acquisition of target information in the image.Compared with support vector machine(SVM),K-Nearest Neighbors(k NN),VGGNet,Goog Le Net and Mobile Net,the method proposed has a better identification accuracy which is able to accurately identify rice pests with small sample size problem under the complex backgrounds.