Scholarlink

<p>Accurate crop disease detection is critical for optimizing input usage and improving yield in precision agriculture. This paper presents KonvLiNA, a modular, anchor-free, single-stage object detection model optimized for crop disease detection. The model introduces an Adaptive KAN Swin Encoder (AKSE) as its backbone, combining Kolmogorov–Arnold Networks (KAN), Swin Transformer blocks, and a gating mechanism for multi-scale feature extraction. For multi-scale feature fusion, KonvLiNA integrates KAN-Enhanced Feature Aggregation in BiFPN (KEFAB), designed for efficient cross-scale information flow. To mitigate information loss during upsampling, an Enhanced Nyström Attention Up-sample (eNAU) module applies Nyström-approximated self-attention to preserve spatial information before fusion. Detection is performed using an anchor-free head with compact KAN layers (cKAN-Head), which predicts classes and bounding boxes regression without predefined anchors. Extensive experiments on several popular datasets shows improvements. For instance, in Rice Leaf Disease detection, KonvLiNA improves average precision (AP) by 4.3% over Faster R-CNN + FPN, 3.5% over YOLOv8, and 0.8% over DINO. Similarly, average recall (AR) increases by 5.7% over Faster R-CNN + FPN, 4.4% over YOLOv8, and 3.8% over DINO, highlighting its robust detection performance across diverse scenarios. Other evaluation includes, PlantVillage, Coconut, Sugarcane, CCMT-Cashew, Cassava, CCMT-Maize, CCMT-Tomato, and COCO for generalization. These results support KonvLiNA as a robust and efficient solution for accurate crop disease detection across varied agricultural conditions <a href="https://github.com/yunusa2k2/KonvLiNA">https://github.com/yunusa2k2/KonvLiNA</a>.</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

KonvLiNA: integrating Kolmogorov-Arnold network with linear Nyström attention for feature fusion in crop disease detection

  • 期刊论文

  • 发表日期：

    2026年1月4日

查看全文
  • Yunusa Haruna,
  • Shiyin Qin,
  • Adamu Lawan,
  • Abdulrahman Hamman Adama Chukkol,
  • Ibrahim Haruna Abdulhamid

摘要

Accurate crop disease detection is critical for optimizing input usage and improving yield in precision agriculture. This paper presents KonvLiNA, a modular, anchor-free, single-stage object detection model optimized for crop disease detection. The model introduces an Adaptive KAN Swin Encoder (AKSE) as its backbone, combining Kolmogorov–Arnold Networks (KAN), Swin Transformer blocks, and a gating mechanism for multi-scale feature extraction. For multi-scale feature fusion, KonvLiNA integrates KAN-Enhanced Feature Aggregation in BiFPN (KEFAB), designed for efficient cross-scale information flow. To mitigate information loss during upsampling, an Enhanced Nyström Attention Up-sample (eNAU) module applies Nyström-approximated self-attention to preserve spatial information before fusion. Detection is performed using an anchor-free head with compact KAN layers (cKAN-Head), which predicts classes and bounding boxes regression without predefined anchors. Extensive experiments on several popular datasets shows improvements. For instance, in Rice Leaf Disease detection, KonvLiNA improves average precision (AP) by 4.3% over Faster R-CNN + FPN, 3.5% over YOLOv8, and 0.8% over DINO. Similarly, average recall (AR) increases by 5.7% over Faster R-CNN + FPN, 4.4% over YOLOv8, and 3.8% over DINO, highlighting its robust detection performance across diverse scenarios. Other evaluation includes, PlantVillage, Coconut, Sugarcane, CCMT-Cashew, Cassava, CCMT-Maize, CCMT-Tomato, and COCO for generalization. These results support KonvLiNA as a robust and efficient solution for accurate crop disease detection across varied agricultural conditions https://github.com/yunusa2k2/KonvLiNA.