Identification Method for Rice Pests with Small Sample Size Problems Combining Deep Learning and Metric Learning

Abstract

AbstractTo achieve accurate identification of rice pests with small sample size problems under complex backgrounds, we proposed a rice pest identification method combining deep learning and metric learning. To overcome the effect of the complex background of the rice pest image, we used the U-Net network which can well retain target information and has a good segmentation effect on rice pests with small sample size problems to remove the background. We also improved the backbone network of VGG16 to extract a more effective feature for identification. Finally, we introduced metric learning to project the feature vector of the pest image to a new feature space for similarity matching and solve the lower accuracy caused by the small sample size. Experimental results showed that the accuracy of the proposed method is better than that of SVM, kNN, AlexNet, VGGNet and Mobilenet. Thus, our method could accurately identify rice pests with small sample size problems under complex rice backgrounds.KeywordsPest identificationSmall sample size problemsDeep learningMetric learningMachine learning