F3M-Det: A Frequency-Guided Three-Modal Framework for Robust Tomato Detection in Complex Agricultural Environments

Abstract

Tomato detection is a fundamental task in intelligent agriculture, yet reliable perception in real production environments remains challenging due to illumination variation, occlusion, and cluttered backgrounds. Although recent detectors have achieved promising performance, many existing methods remain limited in their ability to effectively exploit the complementary appearance, structural, and spectral information available in complex agricultural scenes. To address this issue, we propose F3M-Det, a Frequency-guided Three-modal Mamba Detection Network that integrates RGB, depth, and near-infrared (NIR) modalities through a task-oriented multimodal fusion framework to improve tomato detection robustness. Specifically, the proposed framework introduces a Structural-guided Frequency-aware Mamba Representation Block (SFMRB) to jointly capture geometric structures, spectral characteristics, and long-range contextual dependencies. In addition, an RGB-guided reconstruction fusion strategy is designed to enhance cross-modal consistency and improve feature complementarity. Extensive experiments on a custom multimodal tomato dataset collected under complex agricultural conditions show that F3M-Det achieves 94.26% mAP@0.5, 91.64% mAP@0.5:0.95, 90.82% F1-score, and 91.95% recall, outperforming Faster R-CNN, RT-DETR, multiple YOLO variants, and APNet. Compared with the RGB-only setting, the proposed method improves mAP@0.5 by 5.52 percentage points and mAP@0.5:0.95 by 7.52 percentage points. These results indicate that task-oriented multimodal fusion effectively improves tomato detection accuracy and robustness in challenging agricultural environments.