The Research On Detection Of Torreya Grandis Pests Based On Deep Learning

Torreya grandis belongs to the gymnosperm,Taxus chinensis,and is a unique international rare dried fruit.It is used for fruit,oil,medicinal,ornamental,green,etc.And it also has high economic value.During the growth process,Torreya grandis is extremely vulnerable to a variety of pests.Early warning and forecasting based on continuous monitoring of pest populations is a key component of pest control systems,and plays a major role in information support and decision support in the prevention and control of major pests and diseases.At present,the classification of agricultural pests and the corresponding statistical counts are mainly done manually,with high labor intensity and low work efficiency.The rapid development of deep learning technology in the field of computer vision provides a new way for modern agricultural and forestry automated pest monitoring,which can greatly improve the recognition efficiency,but also solve the problems of lack of agricultural experts and poor objectivity.In order to study the effects of different convolutional neural networks and the effect of transfer learning strategies on the effects of convolutional neural network models based on training in small sample datasets,this paper first explores the performance of various convolutional neural network models in classification tasks based on the small sample datasets of four types of forestry business images including forest pests using transfer Learning.Then a small sample data set of the Torreya grandis pest image was made.Then,based on the complexity of the color texture features of these Torreya grandis pest images with data augmentation,a pest-based automatic detection model based on deep learning was proposed.The main contents and research results of this paper are as follows:(1)This article collects RGB image datasets for 10 species of Torreya grandis pests.Among them,738 individual species of pest images were used as training sets,and 46 images of various pests were mixed as test sets.The data set is manually labeled using Labellmg and saved as a VOC2007 data set.The resolution,blur degree,and brightness of the data in the dataset are different.In order to increase the complexity of the dataset,species don't belong to Torreya grandis pest and different interferents such as dust and broken limbs are added to some images.It can greatly simulate the pest images captured by actual insect traps.(2)In order to analyze the effects of different convolutional neural networks,and the validity and reliability of the transfer learning strategy applied to the model training on small sample data sets,an automatic classification model of convolutional neural network based on migration learning for images containing forest pests and other 3 types of forestry business is designed.On the four models of Inception-v1,Inception-v2,Inception-v3 and VGG-16 which had been proven to have excellent effects in the task of image classification pre-trained by the large-scale auxiliary image dataset ImageNet,the training model is transfered using relatively small forestry business image data.The original 1000-channel fully-connected layer and the softmax layer are replaced with a new fully-connected layer,and then the softmax layer of the four types is added,and the other layer parameters remain unchanged.On the four sets of forestry business image datasets established,the four pre-training convolutional neural network structure transfer learning models have higher classification accuracy.Among them,the migration learning model based on Inception-v3 has the highest recognition accuracy,reaching 96.4%,which proves the effectiveness of the convolutional neural network model based on migration learning on small dataset image classification tasks.Among them,the recognition learning model based on Inception-v3 has the highest recognition accuracy,reaching 96.4%,which proves the validity of the convolutional neural network model based on transfer learning on the image task of small sample dataset.(3)Based on the collected data of the Torreya grandis pests,an automatic detection method of the pests was established by the convolutional neural network.Using the excellent object detection model RetinaNet as the basic model of this paper,by changing the feature extractor enhancing effect,and using the fire module instead of the level with large numbers of channel in the feature pyramid to achieve model compression,and Appling transposition convolution to increase deep feature map resolution and increase receptive field.The transfer learning strategy is adopted.The pest dataset was trained to achieve excellent results in the detection and classification of these pests,with a mean average precision(mAP)of 86.96.It is proved that the automatic detection system using deep learning can play an important role in pest prevention and control.