FSF-RTDETR: a frequency–spatial feature fusion network for eggplant fruit disease and pest detection in complex agricultural scenes
- 1Weifang University of Science and Technology
Abstract
Detecting fruit diseases and pests in complex agricultural scenes remains challenging due to large target-scale variation, weak lesion features, leaf and branch occlusion, and severe background interference. To address these problems, this study focuses on eggplant fruit disease and pest detection and proposes FSF-RTDETR, a frequency–spatial fusion model. The proposed model builds a dual-branch frequency–spatial enhanced backbone to improve disease and pest feature representation from both frequency details and spatial structures. In the frequency branch, AFDC and MFFWBlock enhance high-frequency responses such as fruit borer edges, rot lesion textures, and thrips bite marks through attention-guided frequency-domain convolution and multiscale wavelet feature modeling. This design improves the perception of small targets with weak boundaries and fine textures. In the spatial branch, VisionClueMerge and G6MambaLayer preserve local spatial cues and introduce six directional scanning paths covering horizontal, vertical, and diagonal interactions. This helps the model better describe occluded targets, complex backgrounds, and irregular lesion distributions. Furthermore, EGMuF adaptively integrates edge-texture information from the frequency branch and structural semantic information from the spatial branch through channel, spatial, and pixel-level gated attention. Experimental results show that FSF-RTDETR achieves Precision, Recall, mAP ₅₀ , and mAP 50:95 of 81.5%, 71.3%, 78.8%, and 50.7% on the Eggplant dataset, respectively, outperforming the RT-DETR-R18 baseline by 1.7, 2.0, 4.2, and 4.7 percentage points. The model achieves an inference speed of 104.07 FPS. Results on the PlantDoc Dataset further verify the effectiveness and generalization ability of the proposed method in plant leaf disease detection scenarios.
Related
- Stage-wise collaborative feature enhancement for accurate maize insect pest detection in complex agricultural scenes
- Dual-domain self-supervised feature alignment via spectral–spatial representation learning for deepfake anomaly detection
- Frequency and Laminar Profile of Feature Specific Visual Activity Revealed by Interleaved EEG-fMRI
- Topographic and economic drivers of agricultural–ecological coupling in Gansu Province: spatial differentiation and threshold effects
- Characterizing fatigue-related neural signatures via multi-band EEG features and a three-stage feature optimization strategy for driver fatigue detection
BibTeX
@article{sun2026fsfrtdetr,
title = {{FSF-RTDETR: a frequency–spatial feature fusion network for eggplant fruit disease and pest detection in complex agricultural scenes}},
author = {Hui Sun and Zhijun Zhang and Jun Liu and Delong Zhang and Minghan Feng},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1956953},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1956953/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - FSF-RTDETR: a frequency–spatial feature fusion network for eggplant fruit disease and pest detection in complex agricultural scenes
AU - Hui Sun
AU - Zhijun Zhang
AU - Jun Liu
AU - Delong Zhang
AU - Minghan Feng
PY - 2026
DA - 2026/10/23
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1956953
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1956953/full
AB - Detecting fruit diseases and pests in complex agricultural scenes remains challenging due to large target-scale variation, weak lesion features, leaf and branch occlusion, and severe background interference. To address these problems, this study focuses on eggplant fruit disease and pest detection and proposes FSF-RTDETR, a frequency–spatial fusion model. The proposed model builds a dual-branch frequency–spatial enhanced backbone to improve disease and pest feature representation from both frequency details and spatial structures. In the frequency branch, AFDC and MFFWBlock enhance high-frequency responses such as fruit borer edges, rot lesion textures, and thrips bite marks through attention-guided frequency-domain convolution and multiscale wavelet feature modeling. This design improves the perception of small targets with weak boundaries and fine textures. In the spatial branch, VisionClueMerge and G6MambaLayer preserve local spatial cues and introduce six directional scanning paths covering horizontal, vertical, and diagonal interactions. This helps the model better describe occluded targets, complex backgrounds, and irregular lesion distributions. Furthermore, EGMuF adaptively integrates edge-texture information from the frequency branch and structural semantic information from the spatial branch through channel, spatial, and pixel-level gated attention. Experimental results show that FSF-RTDETR achieves Precision, Recall, mAP ₅₀ , and mAP 50:95 of 81.5%, 71.3%, 78.8%, and 50.7% on the Eggplant dataset, respectively, outperforming the RT-DETR-R18 baseline by 1.7, 2.0, 4.2, and 4.7 percentage points. The model achieves an inference speed of 104.07 FPS. Results on the PlantDoc Dataset further verify the effectiveness and generalization ability of the proposed method in plant leaf disease detection scenarios.
SN - 1664-462X
ER -