Research On Image Recognition Of Peach Pests Based On Deep Learning

Peach trees are susceptible to pests that affect the quality and yield of peaches.It is an important task to identify and prevent and control peach tree pests in a timely and accurate manner.Since there are many kinds of peach pests in natural scenes,it is time-consuming and costly to rely only on the experience of agricultural experts and peach farmers to determine the pest species of peach trees,which is subjective and prone to errors.In recent years,the application of deep learning in agricultural pest image recognition has received a lot of attention from researchers at home and abroad.Deep learning has powerful feature extraction ability,which can automatically learn the basic features and deep semantic features of pest images,providing strong support for fast and accurate recognition of peach tree pest images.In this paper,deep learning-based peach tree pest image recognition is studied,and the main research work is as follows:(1)A peach tree pest image recognition model based on multi-scale attentional residual network is proposed for the problem that the color of peach tree pests is similar to the background and the size difference of pest individuals in natural scenes.The convolutional neural network has fixed size of convolutional kernels in the same layer and equal weights for each channel,which makes it difficult to focus on extracting peach tree pest features.Based on the above problems,the residual network is improved as the backbone network.Firstly,the first layer of large convolution of the residual network is replaced by a multi-scale convolutional layer,which avoids the information loss problem of extracting small-sized pest features with large convolutional kernels.Then a selective kernel convolution module is introduced in the residual unit,which adaptively adjusts the convolution kernel weights to focus on extracting pest region features,generating effective perception and suppressing the background interference problem.Experiments show that the recognition accuracy of the proposed model is 93.27%,which is a significant improvement compared to the recognition effect of standard residual networks.(2)A Conv Ne Xt-based peach tree pest image recognition model is proposed to address the problem of low accuracy of pest image recognition due to large intra-class morphological variation of peach tree pests in natural scenes.Firstly,Conv Ne Xt is used as the backbone network and the model structure is optimized by introducing the visual perceptual field module to extract discriminative features of peach tree pests.Then a joint loss function is used to continuously learn the features that minimize the intra-class distance to accelerate the model convergence and improve the recognition accuracy.Finally,two tandem fully connected layers are used to replace the original fully connected layer,and the feature vectors input to the fully connected layer are visualized to show the pest classification effect.The experimental results show that the proposed model improves the recognition accuracy of peach tree pest images in natural scenes compared with 5 networks such as Efficient Net.(3)A lightweight peach tree pest image recognition model based on global context modeling is proposed for existing pest image recognition models with high structural complexity,large runtime dependence on computing resources and difficulty in deployment.Firstly,a lightweight global context modeling module is embedded in the Shuffle Net V2 unit to aggregate all the location information of the input image,capture long distance dependencies,obtain global context features,correct pest information,and improve the feature representation capability of the model.Then the ordinary convolution layer of Shuffle Net V2 is optimized to add lightweight convolution to enhance the feature extraction ability of the model at different stages.Finally,the Softmax classifier is connected to obtain a lightweight peach tree pest image recognition model.After experimental verification,the proposed model outperforms neural networks such as Alex Net in terms of recognition accuracy,parameter computation and floating point operations,and the number of parameters is only one tenth or even a few tenths of Alex Net,which is a lightweight and efficient neural network model that can be easily deployed.