Research On Small-scale Pest Target Detection Method Based On Deep Learning

During the growth of greenhouse crops,accurately and objectively obtaining the changes in pest species and population numbers is of great significance to the yield and quality of greenhouse crops.In the use of image vision technology to detect thrips and whiteflies that often occur in greenhouses,traditional machine learning has low accuracy and high missed detection rate,while deep learning can automatically learn target feature information from images to improve Detection accuracy.However,the current deep learning technology still has difficulties in detecting small targets,and it is difficult to accurately detect small-sized pests.In order to solve this difficulty,this paper proposes an automatic detection method for small-sized pests such as thrips and whitefly based on deep learning.The main work of this paper includes:(1)The necessity of detection of pests in greenhouses,the current research status and the characteristics of small-scale pests in greenhouses are analyzed.The deep learning target detection method is applied to the detection of small-scale pests,and the YOLOv4 network model is used to detect the small-scale pests thrips and whiteflies on the trap image and analyze the detection results.(2)Aiming at the problem that the adults of thrips and whitefly are difficult to detect due to their small size,a small-sized pest detection method based on image area division is proposed.First,the collected images of thrips and whitefly traps in the greenhouse are divided into multiple sub-images by region,and the sub-images are detected by the YOLOv4 network model.The detection results of the small images are mapped to the original trap image,and then The detection results of small images are combined by non-maximum suppression,and finally the detection of thrips and whiteflies on the trap image is realized.(3)For the YOLOv4 network model,there is a phenomenon of missed detection and poor model robustness when detecting a small image of 416×416 pixels.A new network model based on improved YOLOv4 is proposed for the detection of thrips and whiteflies in greenhouses.First,on the YOLOv4 network structure,the feature map at 8 times downsampling is upsampled by 2 times,and the feature map at 4 times downsampling is spliced together to establish a new YOLO detection layer at 4 times downsampling.Furthermore,in view of the problem that the characteristics of the pests to be detected are not obvious due to the degradation of the image quality of the field traps,the residual unit in the first two residual blocks in CSPDarknet53 is enlarged by 4 times,so as to extract the shallow layer of the target pest more accurately Features and location information.Experimental results show that the improved YOLOv4 network has significantly improved the detection accuracy of thrips and whiteflies,the recall rate has been significantly improved,and the number of missed detections has decreased significantly,which has improved the accuracy of the detection of small-sized pests in the greenhouse.