A density-point network for dense tiny stored grain pest counting

Abstract

Monitoring stored grain pests is essential for food security and loss prevention. Traditional pest counting methods, such as bounding box-based and density map-based approaches, face challenges with tiny pests (<5 mm) and dense distributions. Overlapping annotations in bounding box methods lead to over- or under-counting, while density maps ignore isolated pests in sparse regions. Point-based methods improve on isolated pest detection but lack reliability in dense areas. To address these limitations, we propose the Density-Point Network (DP-Net), which integrates density maps and point regression for robust pest counting. DP-Net employs a backbone network to extract image features, which are processed by a Point-Regression Module for pest coordinates and a Density-Map Generating Module for pest distribution. A patch-select strategy combines these outputs to improve counting accuracy. Our experiments, conducted on a dataset of four stored grain pest species, demonstrate that DP-Net achieves an MAE (Mean Absolute Error) of 3.13 and an MSE (Mean Squared Error) of 5.63, outperforming traditional methods. These findings highlight DP-Net's effectiveness in diverse pest density scenarios, making it a promising solution for automated pest monitoring.