Research On Pest Monitoring System Based On Tiny Object Detection

Field pests and diseases are important factors affecting agricultural yield and agricultural product quality.Monitoring and early warning of their development trend is of great significance to improve agricultural product quality and yield and reduce management cost.Traditional field pest monitoring mainly relies on manual monitoring by grassroots agricultural technicians,which has shortcomings such as high monitoring cost,poor timeliness and poor reliability.In recent years,with the development of deep learning and computer vision,field pest monitoring scheme based on intelligent vision can realize large-area monitoring,automatic detection and identification of pests and diseases,and statistical analysis of pest trends,which has become a research hotspot in the field of smart agriculture.However,when deploying the field pest monitoring system based on computer vision,due to the long distance between the monitored object and the camera,and the fact that some pests themselves are particularly tiny,pests appear as tiny objects(pixels less than 32×32)in the collected images,which leads to a series of problems:(1)The pests in the image are tiny and dense,which makes the detection accuracy of the classical detection model extremely reduced,resulting in missed detection and misdetection of pests and diseases.(2)At present,there are few datasets of tiny pests in agriculture,which is difficult to support the needs of deep learning model training.(3)Previous work did not consider the consistency of geometric features and semantic features of cross-scale feature responses,resulting in the difficulty of locating tiny objects.Therefore,as an important branch of object detection,tiny object detection has important theoretical and research value in the field of smart agriculture.In response to the above questions,this thesis works from the following aspects:1.In this thesis,an adaptive tiny object detection algorithm based on the consistency of geometric and semantic features is proposed to solve the problem of difficult location of tiny objects.This method proposes an instance-level shape and semantic consistency supervision module(Instance-level,Scale-adaptive,Shape-preserved,and Semanticconsistent Supervision,I4S),High Resolution(HR),and Multiscale Fusion module to learn how to locate tiny objects.Among them,the I4 S module is mainly composed of two parts:(1)Cross-scale three-dimensional elliptical cone module;(2)Elliptic conic supervision function based on prior knowledge.The three-dimensional ellipsoid cone is mainly used to model the multi-scale of each instance at the same time,which not only obtains the classification information in the multi-scale feature map,but also learns the shape information of the instance in the continuous scale.The elliptic conic supervised function based on prior knowledge uses Gaussian distribution to drive the predicted response value closer to the true.The HR module mainly includes:(1)Generating highresolution feature maps;(2)Introduced an improved Focal Loss function.The Multiscale Fusion module aims to enhance the feature information of each scale of the feature map.Finally,a large number of experimental results show that the proposed method can improve the detection performance of tiny objects.2.This thesis presents a tiny pest detection algorithm based on adaptive-gating branch.It solves the problems of the disappearance of the feature response of tiny pests in agriculture,the sharp decline in accuracy caused by the influence of Io U in detection results,and the lack of tiny pest images in the existing datasets currently disclosed.The previous algorithm is mainly used to improve the network for common tiny objects in terms of geometric and semantic feature consistency,while this algorithm is used to solve the tiny objects problem of agricultural pests in terms of capturing object features in the shallow layer of the backbone:(1)The algorithm integrates adaptive-gating branches into the backbone and supervises the feature response of tiny objects in the branch.(2)A scaleadaptive weighted detection loss function is proposed.The adaptive weight parameters are used to strengthen the learning of tiny objects by the network,which alleviate the imbalance of feature responses between large objects and tiny objects.(3)A dataset of tiny citrus pests was constructed,and then the images were annotated.At the same time,various data augmentation techniques were used to form a dataset containing 38620 pests.Finally,the proposed method is used to test the dataset,and ablation experiments and comparative experiments are performed on the algorithm,the experimental results show that the proposed method strengthen the characteristic response of tiny pests and improve the detection performance.3.Taking citrus pest detection as the starting point,and integrating the above tiny object detection algorithm,the agricultural pest detection system is designed and realized,which uses HTML,CSS and other technologies to achieve front-end development,Python and other programming languages to achieve back-end development,and Sqlite to realize information storage.The system contains two modules: user information module and pest detection module.It meets the real-time detection task of tiny pests in offline scenarios.