Detection And Recognition Of Rice Pests Based On Multi-scale Feature Enhancement And Feature Fusion

Rice is an important food crop in my country,but it is easy to bear the risk of pests in all stages of its growth,and the output is affected and endangers the country’s food security.The detection,identification and prevention of rice pests is imminent.The traditional detection of rice pests is mainly carried out by manual visual observation,which is highly subjective and difficult to promote,and cannot meet the actual needs of pest detection.With the rapid development of computer vision and deep learning,it provides sufficient support for rice pest detection based on deep learning.This paper studies the problem of rice pests susceptible to complex background interference and detection of small target pests.The main research contents are as follows:(1)In view of the lack of public data sets for rice pests,based on the rice pest data set,this paper proposes a self-made rice pest image data set with a total of 22,315 images of 11 types of pests through data augmentation.,providing a large dataset for follow-up pest research.(2)Aiming at the problem that the YOLOv3 algorithm is not effective for small target recognition and the detection scale of rice pests is different,this paper proposes a multiscale feature fusion target detection algorithm YOLOv3-F.First,a Spatial Pyramid Pooling(SPP)module is added between the backbone and neck to fuse high-level features rich in semantic information with low-level features rich in location information,which improves the detection effect of items of different scales;secondly,when predicting In this stage,a prediction layer with higher resolution and richer location information is added,which is more conducive to the identification of small target items;finally,the loss function is changed to CIo U to improve the efficiency of target detection frame regression and reduce the target missed detection rate.Experiments were carried out on the self-made item dataset and the rice pest dataset,and the recognition accuracy was improved.(3)Aiming at the background interference problem of rice pests,this paper proposes a pest detection algorithm YOLOv5-L that improves the multi-scale fusion and feature attention mechanism of YOLOv5.Replace the feature pyramid(FPN)in the YOLOv5 network structure with a weighted bidirectional(top-down + bottom-up)feature pyramid network(Bi FPN),learn the importance of different input features,and then distinguish different input features.Fusion;adding the CBAM attention mechanism module to the bottom-up feature pyramid helps the network to find areas of interest in images with large regional coverage,suppresses the interference of complex backgrounds,and improves the overall accuracy of pest identification.Experiments were carried out on the home-made rice pest data set and the public pest data set IP102,and the recognition accuracy was improved and the real-time performance was satisfied.(4)In view of the problem that deep learning is rarely used in pest monitoring systems,this paper proposes a rice pest detection system based on deep learning.The system receives the pest pictures,and then transplants the rice pest detection model on the mobile phone through the computer,and finally outputs the pest identification result image.The system includes an image input module,a rice pest monitoring module,a result display module and a pest species module.