Pest Manager: A Systematic Framework for Precise Pest Counting and Identification in Invisible Grain Pile Storage Environment

TL;DR

This paper presents Pest Manager, a comprehensive framework combining a specialized pest detection device, a pest drop dataset, and a multi-task Transformer model to accurately count and identify pests in hidden grain storage environments.

Contribution

It introduces PestMoni, PestSet, and PestFormer, a novel integrated system for precise pest monitoring in invisible grain storage environments, improving accuracy and efficiency.

Findings

PestFormer achieves 99.2% accuracy in pest counting.

PestFormer attains 86.9% accuracy in pest identification.

The framework effectively detects pests beneath grain surfaces.

Abstract

Pest infestations pose a significant risk to both the quality and quantity of stored grain, resulting in substantial economic losses. Accurate and timely pest monitoring is essential, but traditional methods are time-consuming and labor-intensive, and often ineffective at detecting pests beneath the grain surface. In this paper, we introduce a systematic framework: Pest Manager for precise pest counting and identification within the invisible grain pile environment. The framework consists of three components: an improved grain probe trap PestMoni, a pest drop dataset PestSet collected by PestMoni, and a multi-task Transformer-based architecture PestFormer for pest counting and identification. Specifically, PestMoni uses an asymmetric layout of infrared diodes and photodiodes, enabling precise recording of pest drops. Based on PestMoni, we develop PestSet, an infrared-perception dataset for pest drops, including five common grain storage pests. Furthermore, we propose PestFormer, a multi-task model with a conditional modification module to reduce deviations across different PestMonis and thereby normalize the data processing workflow. Extensive experiments validate the design, dataset, and model, with PestFormer achieving state-of-the-art accuracies of 99.2% in counting and 86.9% in identification, highlighting its potential for effective pest management in invisible storage environment.