Scholarlink

Deep learning has revolutionized precision agriculture, particularly in plant disease diagnosis. However, deploying Convolutional Neural Networks (CNNs) on resource-constrained edge devices remains a challenge due to the high computational cost and parameter redundancy of traditional Multi-Layer Perceptrons (MLP) used in classification heads. In this paper, we propose a novel integration of Kolmogorov-Arnold Networks (KAN) into the ResNet50 architecture to address the trade-off between model complexity and classification accuracy for tomato leaf diseases. Specifically, we investigate two architectures: a direct replacement of the MLP head with a Chebyshev KAN layer, and a Dual-Branch network that fuses global context from MLP with fine-grained texture features captured by KAN. Experimental results on the tomato leaf disease dataset demonstrate that our proposed ResNet-KAN model achieves an accuracy of 98.13%, outperforming the ResNet-MLP baseline (97.49%) while reducing the number of trainable parameters by approximately 90% (from 1.05M to 112K). Furthermore, the Dual-Branch architecture achieves a state-of-the-art accuracy of 98.28%. These findings indicate that KANs offer a superior capability in capturing non-linear disease patterns with significantly fewer parameters, making them a promising solution for developing lightweight, high-performance diagnostic systems in smart agriculture.

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

Enhanced Tomato Leaf Disease Diagnosis: A Parameter-Efficient Approach using Hybrid ResNet-KAN Architectures

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hoang-Tu Vo,
  • Nhon Nguyen Thien,
  • Kheo Chau Mui,
  • Huan Lam Le,
  • Phuc Pham Tien

摘要

Deep learning has revolutionized precision agriculture, particularly in plant disease diagnosis. However, deploying Convolutional Neural Networks (CNNs) on resource-constrained edge devices remains a challenge due to the high computational cost and parameter redundancy of traditional Multi-Layer Perceptrons (MLP) used in classification heads. In this paper, we propose a novel integration of Kolmogorov-Arnold Networks (KAN) into the ResNet50 architecture to address the trade-off between model complexity and classification accuracy for tomato leaf diseases. Specifically, we investigate two architectures: a direct replacement of the MLP head with a Chebyshev KAN layer, and a Dual-Branch network that fuses global context from MLP with fine-grained texture features captured by KAN. Experimental results on the tomato leaf disease dataset demonstrate that our proposed ResNet-KAN model achieves an accuracy of 98.13%, outperforming the ResNet-MLP baseline (97.49%) while reducing the number of trainable parameters by approximately 90% (from 1.05M to 112K). Furthermore, the Dual-Branch architecture achieves a state-of-the-art accuracy of 98.28%. These findings indicate that KANs offer a superior capability in capturing non-linear disease patterns with significantly fewer parameters, making them a promising solution for developing lightweight, high-performance diagnostic systems in smart agriculture.