<p>Maize pests and diseases greatly undermine crop yield and quality, but current lightweight detectors still fall short in small-object recognition, edge modeling, and adaptation to field conditions. To bridge these gaps, we construct the Corn-d dataset augmented by a class-balanced Mosaic strategy and design the lightweight detector MSTA-YOLOv11. An improved multi-scale edge refinement and detail enhancement module sharpens edges and textures, thereby boosting the perception of tiny objects. The TK-FocusBlock fuses a Top-<i>k</i> sparse attention with spatial attention, guiding the network to key regions. In addition, the self-designed multi-scale large kernel decomposition attention harnesses multi-branch large kernels and channel attention to capture global semantics and refine multi-scale feature fusion, markedly enhancing robustness in complex environments. Experimental results on Corn-d show that MSTA-YOLOv11 attains 92.3% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\textrm{mAP}_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> and 77.5% <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\textrm{mAP}_{50\text {-}95}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mrow> <mn>50</mn> <mtext>-</mtext> <mn>95</mn> </mrow> </msub> </math></EquationSource> </InlineEquation> with only a marginal increase in computation, delivering gains of 1.4% in <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\textrm{mAP}_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation>, 2.0% in <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\textrm{mAP}_{50\text {-}95}$</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mrow> <mn>50</mn> <mtext>-</mtext> <mn>95</mn> </mrow> </msub> </math></EquationSource> </InlineEquation>, 0.4% in precision, and 2.1% in recall over the YOLOv11n baseline. Compared with mainstream lightweight detectors YOLOv8n, YOLOv11n and YOLO-SDW, the proposed model delivers higher accuracy, sharper boundaries and superior focus under cluttered backgrounds, densely packed small objects and occlusions. The results highlight the strong multi-scale perception and fine-grained discrimination of MSTA-YOLOv11, making it well suited for real-time crop pest and disease monitoring.</p>

MSTA-YOLOv11: an improved detection model for maize pests and diseases under complex scenes

期刊论文
发表日期：

2025年12月8日

查看全文

Lijun Yuan,
Fang Wang

摘要

Maize pests and diseases greatly undermine crop yield and quality, but current lightweight detectors still fall short in small-object recognition, edge modeling, and adaptation to field conditions. To bridge these gaps, we construct the Corn-d dataset augmented by a class-balanced Mosaic strategy and design the lightweight detector MSTA-YOLOv11. An improved multi-scale edge refinement and detail enhancement module sharpens edges and textures, thereby boosting the perception of tiny objects. The TK-FocusBlock fuses a Top-k sparse attention with spatial attention, guiding the network to key regions. In addition, the self-designed multi-scale large kernel decomposition attention harnesses multi-branch large kernels and channel attention to capture global semantics and refine multi-scale feature fusion, markedly enhancing robustness in complex environments. Experimental results on Corn-d show that MSTA-YOLOv11 attains 92.3% $\textrm{mAP}_{50}$ ${mAP}_{50}$ and 77.5% $\textrm{mAP}_{50\text {-}95}$ ${mAP}_{50 - 95}$ with only a marginal increase in computation, delivering gains of 1.4% in $\textrm{mAP}_{50}$ ${mAP}_{50}$ , 2.0% in $\textrm{mAP}_{50\text {-}95}$ ${mAP}_{50 - 95}$ , 0.4% in precision, and 2.1% in recall over the YOLOv11n baseline. Compared with mainstream lightweight detectors YOLOv8n, YOLOv11n and YOLO-SDW, the proposed model delivers higher accuracy, sharper boundaries and superior focus under cluttered backgrounds, densely packed small objects and occlusions. The results highlight the strong multi-scale perception and fine-grained discrimination of MSTA-YOLOv11, making it well suited for real-time crop pest and disease monitoring.