Research On The Image Recognition And Occurrence Estimation Of Stored-Grain Insects

This paper focuses on vertical applications of artificial intelligence techniques in grain storage practices to enhance the intelligence level of stored-grain insect monitoring,aiming for quality preservation and loss reduction.In grain storage,image recognition of insects on the surface of bulk grains and estimation of the occurrence of insects inside grain bulks are critical research issues for industrial applications.Therefore,this paper deals with these two issues in depth.The contributions can be summarized as follows:1.An image dataset named RGB Insect is established to address the scarcity of image data of stored-grain insects.RGBInsect contains 3757 images and over 159,000 labeled insects with various poses,rotation angles,and spatial resolutions.Thus,RGBInsect can supply the research of storedgrain insect image recognition for various visual monitoring scenarios and provide research data for small object detection and fine-grained classification in the computer vision community.For classification and object detection,the performances and problems faced by deep learning methods for stored-grain insect recognition are investigated,and strong baselines are provided.2.A multi-scale insect recognition algorithm is proposed to address the problem of significantly decreasing detection performance with the change of insect scales.A feature pyramid backbone network and anchor generation matching strategy are designed to match different scales of insects,the perceptual fields of different feature layers of the convolutional neural network,and the scale of anchors to effectively extract image features with different spatial resolutions and semantic information.The network expression capability and training effect are improved by adjusting the classification and bounding box regression sub-networks and the restructured loss function.The effectiveness of the proposed algorithm in improving insect recognition accuracy is verified on the RGBInsect dataset.3.The spatial and temporal distribution of insect inside grain bulks is studied based on trapping data to provide a theoretical basis and data support for the research of insect occurrence estimation in grain piles using trapping data.Experiments are implemented to collect insect trapping data,the grain temperature,relative humidity,and manual sampling results at trap locations under multiple storage conditions.Through the statistical analysis,the uneven spatial distribution and the changing pattern over time of trapping data are clarified;the coupling relationship between trapping data and grain temperature and relative humidity at trap locations and the correlation between trapping data and insect densities inside grain bulks are verified;the aggregated distribution of insects inside grain bulks is found in most cases.4.A data fusion-based model for estimating the occurrence of insects inside grain bulks is proposed.Based on the deep belief network and grain isothermal model,this model achieves the feature extraction and data fusion of trapping data,grain temperature,calculated moisture content,and trap location,and realizes the estimation of insect occurrence inside grain bulks at three levels.The validity of the proposed model in estimating insect occurrence is verified through ablative and comparative experiments.The application potential of the proposed model is demonstrated by evaluating the influence of the number of traps on the estimation accuracy and by comparing and analyzing the estimation results of the manual sampling method.