Robust agricultural pest detection under occlusion and environmental variations via AGIIN-MAF training and SAODL...

This study addresses the challenge of pest detection in agriculture, particularly focusing on improving the accuracy and robustness of detection models in varying environmental conditions and when pests are occluded. We propose a novel method, AGIIN-MAF, which includes a main and an auxiliary model. The auxiliary model, equipped with a convolutional autoencoder (CAE) and an Adaptive Gate Information Integration Network (AGIIN), processes occluded images to enhance the main model's ability to detect pests even when they are partially hidden. Our method achieves an mAP50 of 80.2% in detecting occluded objects without increasing the number of model parameters. Furthermore, we introduce a Selective Adaptive Optimization for Domain-aware Learning (SAODL) strategy during the testing phase, which adapts the model to new, changing environments by distinguishing between domain-invariant and domainsensitive parameters. This approach yields a superior average mAP50 of 74.4% across various environmental conditions, outperforming other test-time adaptation methods. Our contributions include enhancing the model's capability to detect occluded pests and improving its adaptability to unseen environments, which is crucial for effective pest management in agriculture.