Research On Pest Detection Method Based On Convolutional Neural Network And Metric Learning

Fast and accurate pest recognition is the key to pest monitoring.Traditional pest recognition process relies on the knowledge of agricultural experts.Manual identification is laborious,time-consuming and error-prone,so that it does not suit the needs of automation in modern agriculture.In recent years,researchers have conducted a lot of research on pest detection methods based on machine vision technology,hoping to build a real-time and high-precision detection system to replace traditional manual methods.Among them,the image processing method based on convolutional neural network surpasses the traditional machine vision method by a large margin in model accuracy,especially in the problem of pest recognition using images captured in the wild.The current pest identification research is mainly based on the classification task of adults,and there are relatively few researches on the identification of single pests that are small in size and similar to its background.In images,these kinds of pests are small in pixel size and similar to the background characteristics.Thus,recognition task of these pests are challenging.On the other hand,for the target detection task of multiple pests,the current research involves multiple digits to dozens of categories,and the capability of the model when there are more pests has not been discussed.At the same time,the feature classification model based on convolutional neural network requires data collection,labeling and re-training when new pest categories need to be added,which is still insufficient in the flexibility and versatility of the model.This thesis will focus on these two issues,the main contributions are as follows:(1)Propsed a convolutional neural network(CNN)based approach to recognize aphid nymphs on leaf surface.The aphid numphs are small sized and having high similarity with their background.The images of pakchoi leaves infested by aphids were collected,and a pixellevel segmentation model of the aphid nymphs in the image was constructed based on the U-Net convolutional neural network model.The trained model achieved a Dice coefficient of 0.82 in the test data.Furthermore,the pixel-level segmentation results are used to count aphids,and the average absolute error between the results and manual counting results is 1.2.Counting precision and recall are both above 95%.Compared with traditional image segmentation methods,the proposed method performed better in both segmentation and counting.(2)Studied insect object detection using the i Naturalist2017 dataset.Designed experiments to validate the cause of low classification accuracy in the i Naturalist insect classification task.From the results,it could be concluded that the problem of different scales of insects in the images is the main factor affecting the classification accuracy,and the interference of the background in the image is a secondary factor.On this basis,a detection method using the Faster R-CNN model to locate insects in the image is proposed,and then the Res Net50 model is used to classify insects after unifying the insect scale.Compared with the model that uses the Faster R-CNN model for end-to-end localization and classification,the proposed model boosted m AP of the i Naturalist insect object detection from0.483 to 0.614.(3)Built a class-agnostic Faster R-CNN model for localizing insects in images using the i Naturalist dataset.In cross-category cross-validation,the model achived an m AP above0.73.On four different agricultural datasets,the model achieved m AP scores above 0.7,without futher fine-tuning using agricultural pest images.(4)Proposed an agricultural pest’s classification model based on convolutional neural network and metric learning.Using the Arc Face metric learning method,a Res Net50 convolutional neural network was trained on the i Naturalist data set as a feature extractor.When applied to classification of agricultural pests,a small number of pest images of known categories are used to extract features and build a feature gallery.During inference of an image,the features are extracted and compared with the feature gallery to obtain the classification results.On four different agricultural pest datasets,classification accuracies of above 90% could be achived using only a small amount of data(5 images per category)to construct the feature gallery.(5)Designed and deployed convolutional neural networks on FPGA hardware.By using lightweighted feature extractors,applying knowledge distillation and leveraging neural network quantization methods,the inference speed of the model can be accelerated while retaining sufficient model accuracy.The test results show that the pest location model runs at a speed of more than 150 images/sec under the premise of an AP50 index of 0.91.The pest classification model runs at a speed of more than 200 images/sec when the accuracy reaches 90%.The aphid nymphs’ segmentation model achieved a Dice coefficient of0.69 when its inference speed is about 3 images/sec.