A hybrid transformer–neuromorphic framework for robust pest detection in agriculture

Abstract Detecting pests in agricultural fields remains one of the challenging tasks due to the small size of the objects and occlusion effects. In this paper, Trans-Neu-Pest, a hybrid framework, has been proposed by integrating the Swin Transformer, an ROI-conditioned NeRF, reinforcement learning-based active vision, and energy-efficient SNN inference. In this proposed model, the improvement of robustness has been applied due to multi-view consistency and adaptive focus mechanisms. Experimental results indicate that the standard datasets demonstrate improved accuracy by reducing false positives and better generalization under different challenging conditions. The proposed framework also introduces the ROI-conditioned NeRF for pest-aware 3D consistency, uncertainty-aware RL for adaptive viewpoint selection, and cross-module interaction between 2D and 3D representations. These modules have been considered as a tightly coupled system where each of the modules works functionally interdependent.