Research On Object Detection Alogorithm Of Stored-Grain Pests Based On Synthetic Images

Effective monitoring of stored-grain pests plays an important role in food security.With the development of deep learning and computer vision technology in recent years,stored-grain pests monitoring technology has gradually changed from traditional manual sampling and counting to image-based automatic monitoring technology.However,the object detection of stored-grain pests is different from the general detection tasks.The object of stored-grain pests is small and there’s subtle difference between different kinds of pests,which makes the detection task difficult;Meanwhile,the pest detection method based on deep learning relies heavily on the annotated image data,and needs to annotate a large number of images artificially;Thirdly,stored-grain pests live in a special environment and move flexibly,which makes data collection and dataset production more difficult and costly.Synthetic image technology is a technology that simulates real image data through a computer.When there is little or no real image data,the computer is used to simulate real data to solve actual computer vision problems.Combining the current situation of stored-grain pest monitoring and the advantages of synthetic image technology,the synthetic image can be effectively and reasonably applied to the problem of stored-grain pest monitoring.The main significance lies in:liberating manpower and reducing the workload of stored-grain pest image collection and dataset production,which greatly reduces the time cost.The synthesized image data has high quality and accurate annotation,which can be effectively applied in the research of stored-grain pest detection algorithms.The main research work of this paper include:Firstly,a method for establishing a synthetic image dataset of storedgrain pests is proposed in this paper based on the synthetic image technology,which includes three steps:establishing a three-dimensional model of stored-grain pests,establishing a dynamic generation equation for the synthetic image of stored-grain pests,and establishing stored-grain pests material transformation neural network.Based on this method,this paper establishes a synthetic image dataset named VirtualInsect.It contains sixteen main stored-grain pests and the number of pest objects is more than one million.The characteristics of the VirtualInsect dataset is analyzed in this paper.Secondly,based on VirtualInsect,this paper designs comparative experiments of stored-grain pests image classification and stored-grains pests object detection.The model trained by the real stored-grain pests image dataset RGBInsect is used as the baseline model,and compare it through multiple models.The results verify the validity and correctness of the method for establishing the synthetic image dataset of stored-grain pests and the VirtualInsect dataset.The algorithm model can greatly reduce the use of real stored-grain pests images with a certain loss of accuracy.Thirdly,a model compression algorithm based on synthetic dataset VirtualInsect is designed for stored-grain pests object detection model in this paper.The efficiency of stored-grain pests object detection model is improved without using real pest image data.This algorithm reduces the model parameters by 85.17%and reduces GPU memory by more than 57.9%,and doubles the model inference speed without loss of the accuracy basically.Finally,a web-based stored grain pest dataset platform is established,which realizes the open source of stored grain pest dataset,and has the functions of dataset management,dataset download,user rights management and so on.