Research On Fast Recognition And Segementation Algorithm Of Crop Pests Based On Deep Learning

As a great agricultural country,the prevention and control of pests and diseases in the agricultural production process has been a problem for a long time.At present,the classification and statistical counting of agricultural diseases and insect pests are mainly done manually,which is inefficient and labor intensive.Although traditional machine learning can achieve intelligence process to some extent,it is too dependent on the selection of features.The acquisition of the above features depends on specific problems and professional knowledge.The emergence of deep learning makes the extraction of image features completely handed over to the neural network to extrac the extracted more comprehensive features,which is more conducive to object recognition and segmentation.Although some progress has been made in the identification and segmentation of pests,there are still many challenges,including the following three aspects: First,most of the current research are aimed at identifying and segmenting certain or specific types of pests,and the applicability of the trained models is poor;second,the real-time performance of pest object detection needs to be further improved;third,the research on pest segmentation based on deep learning needs to be further deepened,and the segmentation efficiency needs to be improved.Based on the above problems,this paper analyzes the current research status of crop pest identification and segmentation at home and abroad.We take 4285 images of common insect pests of 37 crops involving 5 insect orders and damage hundreds of crops species as the research object and carry out related research on pest object detection and segmentation based on deep learning.The main research work of this paper includes the following aspects:(1)Aiming at solving the inefficiency,complex preprocessing,and low recognition accuracy of traditional manual recognition and pattern recognition,we conduct related research based on the single-level object detection model YOLO(You only look once).First,Darknet53 and Efficientnet are used as the backbone feature extraction network of YOLOv3.The results show that the performance of the model is better when Efficientnet is used as the backbone network.The average positioning accuracy and recognition error rate are 98.89% and 1.57% respectively.It takes0.048 s to detect each picture.The size of the anchor frame of the pest data set is also analyzed,and the distribution of the target frame aspect ratio of the data set is obtained through the K-means clustering algorithm.The anchor frame parameters of the coco data set and the anchor frame parameters of the data set obtained by K-means clustering are used to perform experiments on YOLOv3,YOLOv3-Tiny,YOLOv4 and YOLOv4-Tiny.The results show that the network trained with the anchor frame parameters of the coco data set has better average accuracy and average recognition than the results of the network trained by the anchor frame parameters of the experimental data set obtained by the K-means clustering algorithm.Then the YOLOv3,YOLOv4,YOLOv3-Tiny and YOLOv4-Tiny models are tested.The classification and positioning results shows that the overall performance of YOLOv3-Tiny is better than other models.The average positioning accuracy and the average recognition error rate are 98.52% and 1.81% respectively.Meanwhile,it takes only 0.030 s to detect each picture.Finally,the abnormal images detection results in the experiment are analyzed and explained.In the following chapters,the results of the YOLOv3-tiny model will be used to compare with other models in the performance of object recognition and position.(2)In the process of agricultural production,the judgment of the current growth and development status of the pests and the number of pests can help the farmers analyze the pest situation more accurately.But object detection can only classify and locate pest objects.We conduct the research on pest segmentation based on Mask R-CNN,a two-level segmentation model that is currently effective.First,Res Net50,Res Net101,and Mobilenet are used as the backbone feature extraction networks to train the model.The results show that the model which with Res Net50 as the backbone network has the best performance among them.The average accuracy of the positioning and mask evaluation is 95.58% and 91.90% respectively.The average recognition error rate is 8.57%,and it takes 0.16 s to detect each picture.Second,the images which may cause segmentation abnormalities are introduced in details.Then the classification and positioning performance are compared between YOLOv3-Tiny and the Mask R-CNN model with Res Net50 as the backbone.When Io U is 0.5,0.6,and 0.75,the average positioning accuracy of YOLOv3-Tiny is 2.94%,3.99%,and24.92% higher than Mask R-CNN,and the average recognition error rate is 6.76%lower than Mask R-CNN.The time of YOLOv3-Tiny is only 18.75% of Mask R-CNN to process an image.Finally,the abnormalities in the experimental results are analyzed and explained.After removing some of the abnormal images,the models are retrained,and the average accuracy of positioning and masking is increased by 1.60%and 1.06% respectively.In the following chapters,the results obtained by the Mask R-CNN model with Res Net50 as the backbone are compared with other models in terms of classification,positioning and segmentation.(3)Although Mask R-CNN realizes object segmentation of pest images basically,its speed needs to be further improved.We carries out related experiments on the single-stage segmentation model YOLACT and its improved version YOLACT++.The YOLACT(You only look at coefficients)and YOLACT++ models are trained and verified with Res Net50 and Res Net101 as the backbone respectively.The results show that the YOLACT++ model with Res Net101 as the backbone has the best comprehensive performance.When Io U is 0.5,the average accuracy of positioning and masking reaches 95.06% and 93.15% respectively,while the average classification error rate is 12.12%,which takes 0.080 s to process each image.In terms of classification and positioning,by comparing the results of YOLOv3-Tiny,Mask R-CNN,YOLACT and YOLACT++ when Io U is 0.5,0.6,and 0.75,YOLOv3-Tiny has the best detection effect,and its average positioning accuracy and average recognition error rate are 98.52% and 1.81% respectively.It takes only 0.030 s to detect each picture.In terms of segmentation,by comparing the results of the three models of Mask R-CNN,YOLACT and YOLACT++,the YOLACT++ model with Res Net101 as the backbone has the best performance.The average accuracy of the mask branch and the average recognition error rate are 93.15% and 12.12%respectively,it only takes 0.080 s to process each pest image.Finally,the abnormalities in the results are analyzed and explained,and the abnormalities appearing are compared with those in YOLO and Mask R-CNN.In summary,the YOLOv3-Tiny model is selected when it is used for pest object detection.If the shape and quantity of pests are needed for statistical analysis,the YOLACT++ segmentation model with Res Net101 as the backbone comes best.The proposed methods can identify,detect and segment pests quickly and efficiently.While brings great convenience to the farmers’ statistical counting of pests and pest control,they aslo provides a data basis for the prediction and early warning of crop pests in the later period,which has certain reference significance for the solution of pest problems in the development of smart agriculture.