Research On Object Detection Algorithm Of Light-trap Rice Planthopper Based On Deep Learning

Pest monitoring lamp is widely used because it is a kind of physical control method with simple operation.However,there are serious false detections and missed detections in the identification of rice planthoppers in light-trap images using traditional pattern recognition methods.In order to improve the accuracy and recall rate of automatic detection of rice planthoppers in light-trap images,this paper proposes an automatic detection method based on deep learning.The main research contents and results are as follows:(1)Research on different detection algorithms of rice light-trapped planthoppers based on deep learning.object detection algorithms can be divided into one-stage,two-stage and anchor-free algorithms.In the paper,three object detection algorithms(Cascade RCNN,YOLOV4,CornerNet)are selected from the three types of algorithms respectively.They are trained by the original image input and the image block input,and the rice planthopper in the light-trap images is detected.The results show that partition before image input is helpful to solving the missed and false detection problems caused by the too small proportion of planthoppers,but there are some incomplete insect bodies framed out at the segmentation line.CornerNet has a high recall rate,but there are a large number of redundant detection detection boxes,because of which CornerNet can not be directly applied in practice.(2)Research and optimization of rice planthopper detection algorithm based on improved CornerNet.The difficulties of rice planthopper detection lie in:(1)the small size of the planthoppers and their small proportion in the image,resulting in difficulty in extracting effective features and serious missed detection;(2)the brown planthoppers are very similar to the white-backed planthoppers.The morphology of planthoppers is various,and the image brightness and contrast are uncertain,resulting in redundant detection box.In this paper,an overlapping sliding window algorithm is proposed to reduce the missed detections of planthoppers by using the prior knowledge of the size and the circumscribed rectangle of planthopper.A detection box suppression method is proposed to remove redundant detection boxes to improve the accuracy of planthopper detection.(3)Training,testing and analysis of rice planthopper detection algorithms based on deep learning.The same training set was used to train six models,which are Cascade RCNN,YOLOV4,CornerNet combined with overlapping sliding window and detection box suppression.Then these models are used to detect planthoppers,and compared with the the model trained in Chapter 2.The test results show that the addition of overlapping sliding windows and target detection box suppression modules greatly improves the accuracy and recall rate of each model.The improved CornerNet algorithm has the best effect,and the m AP for the detection of two planthoppers can reach 0.939,which can meet actual requirements.