Research On Tree Pest Detection And Recognition Bethod Based On Improved YOLOv5

The construction of forestry ecological environment monitoring is an urgent need for the healthy and sustainable development of forestry ecology,and is the key to the protection of forest resources,the construction of ecological civilization and the improvement of forestry pest control system.Rapid,accurate and effective detection of tree pests can curb the spread of diseases and pests,promote the scientific management of forest diseases and pests,and reduce the degree of harm to forestry production and ecological environment construction.Traditional pest identification methods rely on subjective intuition and group experience,which will consume a lot of human,material and financial resources.In recent years,the emerging deep learning image recognition method represented by YOLOv5 algorithm provides strong technical support for the detection and recognition of tree pests.Therefore,this paper takes tree pests as the research object,adopts the improved YOLOv5 model to solve the problem of detection and identification of tree pests,and mainly carries out the following work:(1)The data set of tree pests was constructed.Through field shooting and data crawler method,2,668 images of 9 kinds of tree pests were collected,including American white moth,pinochroma pinochroma,star Longicorn,marmoton bug,frost sky moth,elm phoenix moth,sallow leaf nail,grass pitysophila and mulberry Longicorn.The tree pests in the images were labeled by labelimg,a labeling tool.The labeled data set was divided into training set and test set according to the ratio of 8∶2 to evaluate the model performance.The data set was expanded by a combination of flipping,rotation,cropping,deformation,scaling,brightness adjustment and noise addition.Finally,the data set containing 9498 pest images was obtained.(2)Build a deep learning-based tree pest target detection network architecture,and compare the identification accuracy,recall rate and average accuracy of three mainstream deep learning target detection algorithms(Faster R-CNN,SSD,YOLOv5)on the tree pest data set.Finally,YOLOv5 was selected as the tree pest target detection algorithm.The original YOLOv5 algorithm was improved to solve the problem that some tree pests in the actual scene were overlapping and fuzzy,resulting in low vermin-identification accuracy in the identification task.(3)Improve the original model of YOLOv5s from the direction of attention mechanism.The attention mechanism learns to ignore information irrelevant to the detection task and find the most useful information by learning from the input image.The attention mechanism is introduced into the BackBone of YOLOv5s model to obtain the channel information and position information of image features and optimize the feature extraction process.Experimental results show that: after the introduction of attention mechanism,the performance of YOLOv5s model is significantly improved,mixed attention is better than channel attention,and the comprehensive results of this chapter show that the model performance is the best after the introduction of CA attention module.(4)The feature fusion FPN was introduced to improve the YOLOv5s model,which solved the problem of model target detection under different sizes,and improved the detection ability of the model to targets of different sizes by changing the connection of the network.The introduction of FPN in the Neck end of YOLOv5s can improve the model performance without increasing the computation amount of the original basic model and integrate more features,attention mechanism and FPN to propose the CABi-YOLOv5s model.The experimental results fully show that this model makes the target detection box of input image more accurate,higher confidence,and better detection effect of different sizes and multiple targets.The average accuracy,accuracy and recall rate of the performance parameters of the improved CABi-YOLOv5s network model are the highest 96.9%,96.1% and 94.8%.It is effectively proved that the model has high prediction accuracy,low detection rate and good multi-objective detection effect,which lays a theoretical foundation for the detection of overlapping areas of tree pests,the improvement of classification accuracy and confidence.