DBS-Net: A lightweight dual-brain synergy network for robust chili pepper disease recognition in complex environments

Abstract

Accurate chili pepper disease diagnosis is critical for global crop productivity but is hindered by complex field backgrounds and the computational constraints of edge devices. Current lightweight models suffer from “feature dilution” and fail to balance efficiency and feature representation. Inspired by the dual-pathway mechanism of the biological visual system, this study proposes DBS-Net, a lightweight robust framework that decouples texture perception and semantic understanding for resource-constrained agricultural deployment. DBS-Net integrates three key innovations: (i) a Texture Brain with an Adaptive Channel Reweighting Ghost (ACR-Ghost) module for dynamic noise filtering and fine-grained lesion texture enhancement via statistical-semantic joint modeling; (ii) a Semantic Brain that enables adaptive feature reflux and long-range spatial dependency capture using lightweight Multi-Scale Feature Aggregation (L-MSFA) and LiteCoordAtt modules; and (iii) an optimized dual-brain architecture that reduces the parameter count to 0.98 M (


$$\approx $$

≈



75% fewer than GhostNet-1.0) with a compact 3.75 MB model size and high computational efficiency (207.13 M FLOPs). On the iFLYTEK Chili Disease Dataset (ICDD), a challenging in-field benchmark, DBS-Net attains a mean Top-1 accuracy of 66.75% over three runs (best single run 67.02%), exceeding the GhostNet-1.0 baseline by 8.59 percentage points while using only about one-quarter of its parameters and ranking first among representative lightweight and state-of-the-art models. To verify that these gains generalize beyond a single data distribution, all models are additionally retrained from scratch on an independent public chilli-leaf disease dataset, where DBS-Net again achieves the highest accuracy (98.40%), confirming the robustness and transferability of the proposed design. As the smallest model among the compared networks, DBS-Net mitigates “attention drift” in complex backgrounds and achieves a favorable accuracy–efficiency trade-off suitable for mobile edge deployment, providing a scalable solution for field-deployable chili pepper pest and disease monitoring.