Pest Identification And Counting Of Sticky Image Based On Deep Learning

For a long time,China has regarded agriculture as the foundation for the development of the national economy.In rec ent years,affected by factors such as climate change and crop adjustment,there are more and more crop pests and diseases,the area is more and more widespread,and the degree of disaster is becoming more and more serious.Excessive use of pesticides will cause damage to the environment.And the quality and safety of agricultural products is difficult to guarantee.Sound pest monitoring system can reduce the use of pesticides and improve the level of prevention and control of crop diseases and insect pests.Pest monitoring is an important part of the pest monitoring system,and pest identification and counting is the basis of pest monitoring.Traditional human eye observation and judgment,low efficiency,high labor costs,and sound measurement method,infr ared sensor monitoring and other methods have high requirements on equipment,and are not suitable for agricultural pest monitoring.With the development of computer technology,image-based pest identification and counting methods have become a research hotspot.Most of the methods currently used are artificially designed to extract features,and then combined with traditional machine learning algorithms for identification.This method of artificial design features is subjectively restricted.The introduction of deep learning,by virtue of its automatic extraction of features,greatly promoted the development of machine learning.In this paper,deep learning is introduced into the field of crop pest identification and enumeration.The pest identification of worm plate image is realized by improving the existing target detection framework,and the classification count is realized on the basis of recognition.The main research contents of this paper are as follows:(1)Due to the low resolution of the image of the insect-infested pests collected in this paper,and the small size of the pests contained in the image,the shape and texture structure is more complicated.The ordinary camera is limited by factors such as pixels,focal length and illumination,which will affect the recognition and counting effects.In this paper,the deep learning ESPCN super-resolution reconstruction model is used to super-resolution reconstruction of pest images to reflect more detailed information and improve the accuracy of recognition and counting.(2)The pests on the image of the visceral plate are small in size,and the existing target detection framework is not good for small target detection.In this paper,based on the Faster-RCNN target detection framework,the TDM structure is used to realize the fusion of high-level semantic information features and low-level detail features to improve the detection ability of the target detection model for small targets.The Soft-NMS algorithm is used to replace the original The NMS algorithm to improve the accuracy of partial sticky pest detection.(3)Due to the slow detection speed of Faster-RCNN,real-time detection cannot be realized,and the requirements of the pest monitoring system cannot be met.This paper introduces the real-time SSD target detection framework,and on the basis of SSD,it improves the characteristics of small pests: using deconvolution to achieve high-level and low-level feature fusion,and then using the merged features to reconstruct the feature pyramid,the reconstructed feature pyramid Layer-by-layer testing is performed to train a model suitable for pest identification,and the identification and counting of pests is realized on the basis of identification.