Research On Stored-grain Pests Detection Based On Attention Mechanism

Grain is an important material for a country’s economic development and people’s lives.Due to factors such as temperature and humidity,pests can grow in stored grains.Stored-grain pests not only cause grain losses,but also reduce the quality of grain due to their excreta,and spread diseases,which can harm human health.During grain storage,automatic detection of stored-grain pests is of great significance to timely discover and control their reproduction and invasion,ensuring the safety of grain storage.In this paper,we aim to improve the accuracy of stored-grain pest detection by exploring the effect of different attention mechanisms on the stored-grain pest detection model.A multi-head dynamic attention graph module is proposed to enhance the feature extraction capability of the model and improve the performance of the stored-grain pest detection.The SIoU is proposed as a localization loss to speed up convergence and improve the accuracy of the stored-grain pest detection.The main contributions of this paper could be listed as follows.1.After sorting the collected stored-grain pests of six common types,this paper established two standard datasets for the stored-grain pests detection,i.e.a single-class pest dataset and a multi-class mixed pest dataset.We then collaborated with relevant experts to annotate the data to ensure the accuracy of annotations and facilitate subsequent verification of the associated algorithms.2.This paper explores the impact of different attention mechanisms on stored-grain pest detection and demonstrates that using the attention mechanisms to enhance the feature extraction capability can improve the accuracy of stored-grain pest detection.3.In this paper,a novel multi-head dynamic attention graph module is proposed to establish the long-range dependency of features and dynamically select the adjacent nodes with the highest correlation.This module enhances the feature extraction capability of the stored-grain pest detection model.4.This paper proposes to employ SIoU as localization loss.SIoU introduces angle loss to measure the angle between the predicted box and the real box,and also takes into account the loss of distance,shape,and overlap area between the predicted box and the real box,so as to accelerate the convergence speed and improve the accuracy of the stored-grain pest detection.5.Extensive generalization experiments and analyses have been conducted in this paper to demonstrate the effectiveness and excellent performance of the proposed method in many aspects.The proposed attention mechanism-based method proposed in this paper can accurately detect the six types of grain pests mentioned above,achieving a mean average precision of 97.2%and 91.0%on the singleclass pest dataset and the multi-class mixed pest dataset,respectively.