Lightweight real-time citrus fruit disease detection via multi-scale feature fusion and sparse interactive attention

To address the challenges posed by small disease targets, severe background interference, and high inter-class similarity in real orchard environments, a lightweight and accurate citrus fruit disease detection method, termed MDF-DETR, is proposed. First, a Multi-Scale Selective Feature Fusion (MSSF) module is designed to model lesions of different sizes through gated weighted fusion and multi-granularity parallel large-kernel branches, thereby compensating for the attenuation of high-frequency information from small lesions caused by successive convolution and pooling operations. Subsequently, a Dual-Branch Cross-Scale Adaptive (DBCA) feature fusion network is developed to progressively integrate shallow high-resolution spatial features with deep semantic representations through hierarchical fusion and adaptive weighting. This design preserves lesion boundaries while suppressing interference from complex background noise. Finally, a Focused Sparse Interactive Attention (FSIA) module is introduced to enhance subtle lesion features with reduced computational cost. Through density-driven token aggregation, FSIA preserves full-resolution Query features while compressing Key and Value tokens, reducing self-attention complexity from quadratic to linear under a fixed aggregation limit. This design further strengthens attention to lesion boundaries and texture variations while suppressing background interference. Experiments on the self-constructed JXDF dataset and the public OFDD dataset show that MDF-DETR achieves mAP values of 98.52 ± 0.0018% and 96.54 ± 0.0027%, respectively, representing improvements of 5.35 and 3.67 percentage points over the baseline. With 17.80M parameters and an inference speed of 102.75 FPS, MDF-DETR achieves a favorable balance between detection accuracy and computational efficiency. These results demonstrate the effectiveness and stability of the proposed method for fruit-level citrus disease detection in complex orchard environments and provide a promising basis for subsequent practical deployment.

Authors

Institutions

Publication Details

Journal
Frontiers in Plant Science
Published
2026-11-18
DOI
https://doi.org/10.3389/fpls.2026.2012625
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Lightweight real-time citrus fruit disease detection via multi-scale feature fusion and sparse interactive attention

Baijing Wu, Ke Gao, Guanghui Yan
7 links
Frontiers in Plant Science
Smart Agriculture and AI
article

Lightweight real-time citrus fruit disease detection via multi-scale feature fusion and sparse interactive attention

Baijing Wu, Ke Gao, Guanghui Yan
article en

Abstract

To address the challenges posed by small disease targets, severe background interference, and high inter-class similarity in real orchard environments, a lightweight and accurate citrus fruit disease detection method, termed MDF-DETR, is proposed. First, a Multi-Scale Selective Feature Fusion (MSSF) module is designed to model lesions of different sizes through gated weighted fusion and multi-granularity parallel large-kernel branches, thereby compensating for the attenuation of high-frequency information from small lesions caused by successive convolution and pooling operations. Subsequently, a Dual-Branch Cross-Scale Adaptive (DBCA) feature fusion network is developed to progressively integrate shallow high-resolution spatial features with deep semantic representations through hierarchical fusion and adaptive weighting. This design preserves lesion boundaries while suppressing interference from complex background noise. Finally, a Focused Sparse Interactive Attention (FSIA) module is introduced to enhance subtle lesion features with reduced computational cost. Through density-driven token aggregation, FSIA preserves full-resolution Query features while compressing Key and Value tokens, reducing self-attention complexity from quadratic to linear under a fixed aggregation limit. This design further strengthens attention to lesion boundaries and texture variations while suppressing background interference. Experiments on the self-constructed JXDF dataset and the public OFDD dataset show that MDF-DETR achieves mAP values of 98.52 ± 0.0018% and 96.54 ± 0.0027%, respectively, representing improvements of 5.35 and 3.67 percentage points over the baseline. With 17.80M parameters and an inference speed of 102.75 FPS, MDF-DETR achieves a favorable balance between detection accuracy and computational efficiency. These results demonstrate the effectiveness and stability of the proposed method for fruit-level citrus disease detection in complex orchard environments and provide a promising basis for subsequent practical deployment.

Frontiers in Plant ScienceVol. 17
Lanzhou Jiaotong University (CN), Nanchang Vocational University (CN)
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon