Study And Improvement On Acquisition And Recognition System Of Pest Image In Cotton Field Based On Machine Vision

How to identify pests efficiently and accurately is of great significance for effective pest control at present.Currently,most pests are artificially identified by plant protection experts in traditional manners based on their planting experience.Those time-consuming and arduous manners are featured by their low identification accuracy,large error and instability.Therefore,it is urgent to develop an efficient and precise image acquisition and recognition system against cotton pests.The original image acquisition and recognition system against cotton pests was investigated and improved based on the previous efforts made by the team.The main research are as follows:(1)A field experiment was conducted through the original device for image acquisition against cotton pests in light of their phototaxis.According to the experiment,the images from the device showed issues of excessive and inadequate background removal and inability to identify multi-target pests.In this regard,a research scheme was determined to improve the original device’s structure and to verify the improvement through object detection model.(2)The images of pest samples collected from cotton field were taken in the laboratory through a simple image acquisition device.The data set CPest Data-1 with 3,000 experimental images was constructed.Then,the improved device was adopted to conduct a field experiment.A total of 1,200 field images were used,which were combined with experimental image data set to construct CPest Data-2,a mixed data set.Label Img was adopted to label the above data set manually.The label covers the information on category and coordinates of target pests.The Mosaic data enhancement was adopted to expand the data set and enrich the background of objects detected during the process of model training so as to improve the robustness of the model.(3)Comparison and contrast were made via the constructed data set CPest Data-1 between two types of object detection algorithms based on two-stage Faster R-CNN and one-stage YOLO v4 respectively.The results showed that the m AP of Faster R-CNN was 99.40%,which was only 0.44% higher than that of YOLO v4.However,the detection speed of Faster R-CNN was 0.26 s in each image while the detection speed of YOLO v4 was 0.03 s,which was 8.39 times of that of Faster R-CNN.Eventually,YOLO v4 object detection algorithm was taken as the pest identification model herein in light of the real-time detection needs in the field.(4)The information on coordinates in CPest Data2 data set was clustered through K-Means algorithm,where the optimal distribution parameters of anchors were determined and adopted for YOLO v4 parameter setting.Three attention mechanism modules,namely SE-NET,CBAM and ECA,were adopted to improve the branch of YOLO V4 network structure in sequence.The m APs of SE-YOLO v4,CB-YOLO v4 and ECYOLO v4 models were improved by 0.39%,2.53% and 1.8%,respectively,as compared with that of original YOLO v4 algorithm.Among them,the m AP of CB-YOLO v4 was 94.48%,higher than that of other models.Finally,CB-YOLO v4 object detection model was taken for the subsequent cotton pest identification.(5)An improved image acquisition and recognition system was designed and constructed against cotton pests based on CB-YOLO v4 model.The image acquiring,positioning and recognizing modules were integrated.500 field images were used to validate the reliability of the improved detection system.The resultant m AP was 74.15% and the resultant detection speed was 0.036 s.