Scholarlink

<p>To overcome the constraints of standard methods for detecting forest pests, especially in complex ecosystems and with small targets, this study presents KFBI-DETR, an enhanced model based on Deformable-DETR. The framework incorporates three specialized modules: Convolutional KANs attention-enhanced block (CKAEB), a feature enhancement block leveraging dynamic receptive field adjustment via KAN convolution; fast Fourier transform feature mixer block (FFMB), a frequency spatial fusion module utilizing FFT and partial convolution; and depthwise separable convolution-based upsample block (DCUB), an upsampling module designed with depthwise convolution and channel interaction. Results of experiments on the SE-FPD and FPD datasets indicate that KFBI-DETR outperforms Deformable-DETR, yielding enhancements of 4.2% and 4.3% in mAP, respectively. Notably, it significantly boosts small-object detection accuracy, increasing accuracy from 46.1 to 52.4% on SE-FPD and from 48.7 to 54.9% on FPD. These results surpass other DETR-based variants, which generally yield only marginal gains. These findings highlight the model's effectiveness and practicality in forestry pest surveillance. The code is available at <a href="https://github.com/wanlangyi/KFBI-DETR.git">https://github.com/wanlangyi/KFBI-DETR.git</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

Enhanced forestry pest detection via multi-frequency feature integration and KAN convolution

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Yinglai Huang,
  • Langyi Wan,
  • Yang He,
  • Xueshan Xiong

摘要

To overcome the constraints of standard methods for detecting forest pests, especially in complex ecosystems and with small targets, this study presents KFBI-DETR, an enhanced model based on Deformable-DETR. The framework incorporates three specialized modules: Convolutional KANs attention-enhanced block (CKAEB), a feature enhancement block leveraging dynamic receptive field adjustment via KAN convolution; fast Fourier transform feature mixer block (FFMB), a frequency spatial fusion module utilizing FFT and partial convolution; and depthwise separable convolution-based upsample block (DCUB), an upsampling module designed with depthwise convolution and channel interaction. Results of experiments on the SE-FPD and FPD datasets indicate that KFBI-DETR outperforms Deformable-DETR, yielding enhancements of 4.2% and 4.3% in mAP, respectively. Notably, it significantly boosts small-object detection accuracy, increasing accuracy from 46.1 to 52.4% on SE-FPD and from 48.7 to 54.9% on FPD. These results surpass other DETR-based variants, which generally yield only marginal gains. These findings highlight the model's effectiveness and practicality in forestry pest surveillance. The code is available at https://github.com/wanlangyi/KFBI-DETR.git