A progressive local attention mechanism for fine-grained crop pest identification

Crop pests remain a major factor limiting global crop yields. As such, it is essential to reduce their infestation by identifying individual species and implementing precise control measures. The complex backgrounds occurring in natural scenes and the visually similar species continue to pose unique challenges for large-scale crop pest identification. To address these issues, a novel progressive local attention mechanism (PLAM) is introduced in this paper, which highlights the overall pest structure and acquires discriminative regions containing fine-grained features of the pest through a progressive localization strategy. The method first introduces a correlation energy-oriented mobile inverted bottleneck convolution (CEM) block stacked in a backbone network, a module that implicitly locates the approximate position of the pest while reducing model complexity. Next, a map-derived location and proposal generation (MLPG) module is designed to further reduce interference from complex backgrounds and subtle differences between pests. This module consists of two sub-modules: the activation map-derived location (AMDL) module dynamically locates the region of interest using a global representation, and the feature map-derived proposal (FMDP) module extracts fine-grained features to address the problem of inter-class similarity. The proposed modules were carefully evaluated using ablation studies to verify their effectiveness. In addition, we compared PLAM with representative attention mechanisms and fine-grained visual classification (FGVC) models on the IP102 dataset. The results indicate that the proposed method achieves an accuracy of 89.5%, which outperforms other mainstream methods in recall (+0.7%), F1 score (+1.7%), and accuracy (+1.2%) with lowest model complexity and computational complexity, demonstrating the effectiveness of our PLAM method for fine-grained crop pest identification. The dataset and codes of this paper are available at http://github.com/huanghsheng/PLAM . • PLAM for fine-grained crop pest identification is proposed. • CEM block implicitly reduces the interference from complex backgrounds. • MLPG module minimizes background interference and locates the discriminative regions.