Detection and Identification of Stored-Grain Insects Using Deep Learning: A More Effective Neural Network

Zhichao Shi,Hao Dang,Xiaoguang Zhou,Zhicai Liu

doi:10.1109/access.2020.3021830

Zhichao Shi, Hao Dang + Show 2 more

Open Access

https://doi.org/10.1109/access.2020.3021830

Copy DOI

Abstract

The detection and identification of stored grain insects is important to ensure the safety of grain during grain storage. At present, insect identification methods primarily rely on manual classification; therefore, the automatic, rapid and accurate detection of stored grain insects remains a challenge. This paper proposes an improved detection neural network architecture based on R-FCN to solve the problem of detection and classification of eight common stored grain insects. In this network, we use the multiscale training strategy with a fully convolutional network to extract more features of the insects and automatically provide the location of potentially stored grain insects through an RPN from the feature map. By using the position-sensitive score map to replace some fully-connected layers, our network is more adaptive to detect insects in complicated backgrounds, and our detection speed is improved. In addition, we also used soft-NMS to solve the superposition interference between insects and to further improve the detection accuracy. Sufficient comparative experiments are performed using our two stored grain insect detection datasets, which are carefully annotated by entomologists. Quantitative comparisons against several prior state-of-the-art methods demonstrate the superiority of our approach. Experimental results show that the proposed method achieves a higher accuracy and is faster than the state-of-the-art insect image classification algorithms.

Highlights

Insects are one of the direct causes of loss during postharvest operations [1]; it is crucial to detect and identify stored grain insects by using a stored grain insect monitoring system.Due to the disadvantages of traditional methods [2], such as difficulty in sampling, slow speed, and manual work, rapid and accurate pest detection has long been a difficult problem to solve
We developed a method based on R-FCN, which can be used to detect the insects rapidly and accurately
We proposed the R-FCN+ + + which used soft-non-maximum suppression (NMS) to solve the exciting problem of overlap between stored grain insects

Summary

Introduction

Insects are one of the direct causes of loss during postharvest operations [1]; it is crucial to detect and identify stored grain insects by using a stored grain insect monitoring system.Due to the disadvantages of traditional methods [2] (e.g., near -infrared, acoustical methods, electrical conductivity), such as difficulty in sampling, slow speed, and manual work, rapid and accurate pest detection has long been a difficult problem to solve. The image recognition method based on deep learning can acquire many images of grain insects with low cost and a high recognition rate through seduction and other methods [3]. Deep learning has made many improvements in agriculture through the progress of science and research, such as leaf diseases identification [4] and insect recognition [5], [6]. Deep learning is used to classify the feature. The associate editor coordinating the review of this manuscript and approving it for publication was Shuping He. vectors from insect image features based on the generalized learning ability of big data to quickly identify the categories of different insects. Image detection has notably high research value in the field of stored grain pest detection

Methods

Findings

Discussion

Conclusion