Scholarlink

<p>In smart agriculture, accurate segmentation of maize-leaf diseases in real field imagery supports timely intervention, but remains challenging under cluttered backgrounds, uneven illumination, occlusion, and diverse lesion morphology. We present LKCAFormer, a lightweight encoder–decoder segmentation network that integrates two key components: (i) a three-stage Large-Kernel Cooperative Attention encoder (LK-COAT) that progressively enlarges the effective receptive field via large-kernel depthwise convolutions while preserving fine boundaries using cooperative channel–spatial gating; and (ii) a cross-scale decoder (CSDecoder) that fuses shallow edge/detail cues with deep semantics to refine lesion boundaries at low computational cost. We evaluate LKCAFormer on <i>CD&amp;S</i> and a controlled single-leaf variant derived from it (<i>Single-CD&amp;S</i>), using disease IoU as the primary endpoint. Robustness is further assessed on a 266-image complex-case subset curated from held-out test data, together with paired two-sided tests. On <i>Single-CD&amp;S</i>, LKCAFormer achieves 76.23 ± 2.25 disease IoU and 86.70 ± 1.96 Dice, yielding a modest + 0.58 IoU gain over the strongest lightweight baseline (SwiftFormer). On the more challenging <i>CD&amp;S</i> benchmark, LKCAFormer reaches 69.09 ± 1.65 disease IoU and 78.87 ± 2.13 Dice, outperforming the strongest baseline (SegFormer) by + 4.05 IoU; gains on the complex-case subset are statistically significant. LKCAFormer remains compact (3.68&#xa0;M parameters; 1.13G FLOPs), corresponding to approximately 12.7% of U-Net’s parameters and 1.47% of its FLOPs, while retaining practical end-to-end throughput under a unified profiling protocol. Limitations include fixed dataset splits, the lack of cross-device latency/energy benchmarking, and the absence of multi-seed variability analysis. Future work will extend validation across crops and sensors and provide deployment-oriented, hardware-aware latency and energy evaluations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LKCAFormer: a lightweight transformer with large-kernel cooperative attention for the segmentation of field maize leaf diseases

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月28日

查看全文
  • Jian Hu,
  • Xinhua Jiang,
  • Julin Gao,
  • Xiaofang Yu,
  • XueLiang Fu,
  • Chengjun Zhai

摘要

In smart agriculture, accurate segmentation of maize-leaf diseases in real field imagery supports timely intervention, but remains challenging under cluttered backgrounds, uneven illumination, occlusion, and diverse lesion morphology. We present LKCAFormer, a lightweight encoder–decoder segmentation network that integrates two key components: (i) a three-stage Large-Kernel Cooperative Attention encoder (LK-COAT) that progressively enlarges the effective receptive field via large-kernel depthwise convolutions while preserving fine boundaries using cooperative channel–spatial gating; and (ii) a cross-scale decoder (CSDecoder) that fuses shallow edge/detail cues with deep semantics to refine lesion boundaries at low computational cost. We evaluate LKCAFormer on CD&S and a controlled single-leaf variant derived from it (Single-CD&S), using disease IoU as the primary endpoint. Robustness is further assessed on a 266-image complex-case subset curated from held-out test data, together with paired two-sided tests. On Single-CD&S, LKCAFormer achieves 76.23 ± 2.25 disease IoU and 86.70 ± 1.96 Dice, yielding a modest + 0.58 IoU gain over the strongest lightweight baseline (SwiftFormer). On the more challenging CD&S benchmark, LKCAFormer reaches 69.09 ± 1.65 disease IoU and 78.87 ± 2.13 Dice, outperforming the strongest baseline (SegFormer) by + 4.05 IoU; gains on the complex-case subset are statistically significant. LKCAFormer remains compact (3.68 M parameters; 1.13G FLOPs), corresponding to approximately 12.7% of U-Net’s parameters and 1.47% of its FLOPs, while retaining practical end-to-end throughput under a unified profiling protocol. Limitations include fixed dataset splits, the lack of cross-device latency/energy benchmarking, and the absence of multi-seed variability analysis. Future work will extend validation across crops and sensors and provide deployment-oriented, hardware-aware latency and energy evaluations.