Research On Insect Lightweight Detection Model In Natural Background Based On Deep Learning

Accurate and timely detection of insect infestation is the key technology of monitoring insect infestation in agriculture and forestry.Traditional pest detection mainly relies on a manual method to identify and count pests,which is laborious,time-consuming and errorprone,and difficult to meet the practical application requirements.In recent years,researchers at home and abroad have carried out a large number of studies on pest detection methods.Among them,image processing methods based on deep learning greatly surpass traditional machine vision methods in model accuracy and generalization ability,showing strong robustness in pest image detection,but there are still the following problems: The detection accuracy of pests with small size and similar background in complex natural environment needs to be improved.Due to the high complexity of the model,large number of parameters and high computing overhead,it is difficult to deploy the model on the terminal device or the real-time detection after deployment is poor.The purpose of this study is to improve the detection accuracy and speed of the insect detection algorithms,and combine the model compression technology to make the detection model more lightweight.This paper improves the YOLOv5 s algorithm,and the main work is as follows:(1)An improved YOLOv5 s insect detection model was proposed to solve the problem of mis detection and missing detection of densely stacked small target insects in the natural environment.Firstly,a channel attention mechanism was embedded in the trunk network to improve the feature extraction ability of the algorithm for small target pests.Then,the adaptive spatial feature fusion structure ASFF was introduced in PANet,and the dynamic weight parameters were used to assign different weights to the feature maps of different scales to filter out the features of other levels such as the complex environment of the pests.Finally,the loss function and the non-maximum suppression strategy are changed to improve the speed and confidence of boundary frame localization.Comparing the improved insect detection algorithm with other mainstream target detection algorithms,the experimental results show that the improved algorithm has the best detection effect on the insect data set of this study.The average accuracy of m AP can reach 97.8%,and the average detection time of a single image is13.66 ms.The detection accuracy is greatly improved while the detection speed is guaranteed.(2)Aiming at the problem of large parameter redundancy in the improved YOLOv5 s insect detection model,a model compression and lightweight method based on channel pruning was proposed.Firstly,LI regularized sparsity training is performed on the BN layer of the network,and the trend scaling factor tends to 0 to increase the sparsity of network channels.Then,on the basis of channel pruning,the least square method was introduced to reconstruct the output error before and after the convolution layer to minimize the output error,and the pruning threshold was solved and the pruning was completed at one time.Finally,fine-tuning experiments are carried out to restore the accuracy of the algorithm temporarily lost by network channel pruning.The experimental results show that the memory occupied by the model is reduced by half,the number of parameters is reduced to 60.3%,flop is reduced by 36.8%,and the inference speed is increased by 6.07 ms.Without losing the original precision,the computation and model space of the network are greatly reduced,and the network inference speed is accelerated.All experiments in this study were conducted on a publicly available insect image data set.The experimental results show that the proposed insects lightweight detection under complicated background of natural environment model,implement the detection accuracy and speed of good balance and ascension,for landing deployment insect detection model provides an effective solution.