Scholarlink

<p>With the development of smart agriculture and the growth of the global population, vegetable production is facing the dual challenges of diversified planting environments and increased concealment of diseases. Eggplant, as an important economic crop, has its disease detection accuracy directly affecting yield and quality. However, traditional detection methods fail to effectively capture small diseased areas. To address this issue, this paper proposes an improved deep learning small target detection model—the Efficient Deep-fusion Detection Model (EDDet), which is specifically optimized for the recognition of small diseased spots in eggplant disease detection. In the detection network, we innovatively designed the Pinwheel Fusion Feature Extractor (PFFE) framework, replacing the standard convolutions of the first two layers with Pinwheel Convolutions (PConv). By using asymmetric padding and parallel convolution kernel design, the receptive field is effectively expanded, the ability to capture underlying features is enhanced, and the detection of small diseased areas in eggplants is more precise. In the feature fusion stage, this paper designs a Cross-layer Attention Module (CAM), including Cross-layer Channel Attention (CCA) and Cross-layer Spatial Attention (CSA), which can efficiently interact and fuse features of different scales without additional sampling, alleviating the information loss caused by semantic gaps. In addition, to solve the instability caused by IoU fluctuations in the bounding box regression process, the model introduces Scale-based Dynamic Loss (SD Loss), which dynamically adjusts the loss weight based on the size of the target. By adaptively adjusting the proportion of IoU-based loss and location constraint loss, more precise localization and stable regression of small diseased areas in eggplants are achieved. Experimental results demonstrate that EDDet achieves a notable improvement in mAP50 (85.4%), outperforming the baseline by 2.8%.Importantly, EDDet also Maintains excellent efficiency with only 2.75&#xa0;M parameters, 9.1 GFLOPs, and a high inference speed of 288.3 FPS, which is 37.5 FPS higher than the baseline.These results highlight the model’s strong potential for real-time deployment in complex agricultural scenarios where both precision and speed are critical.</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

EDDet: efficient deep-fusion and dynamic optimization for small target detection in eggplant diseases

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年9月30日

查看全文
  • Ye Li,
  • Xiaofang Li,
  • Rui Fu,
  • Zhenqi Cheng,
  • Jianchun Yu,
  • Xuewei Wang,
  • Hao Sun

摘要

With the development of smart agriculture and the growth of the global population, vegetable production is facing the dual challenges of diversified planting environments and increased concealment of diseases. Eggplant, as an important economic crop, has its disease detection accuracy directly affecting yield and quality. However, traditional detection methods fail to effectively capture small diseased areas. To address this issue, this paper proposes an improved deep learning small target detection model—the Efficient Deep-fusion Detection Model (EDDet), which is specifically optimized for the recognition of small diseased spots in eggplant disease detection. In the detection network, we innovatively designed the Pinwheel Fusion Feature Extractor (PFFE) framework, replacing the standard convolutions of the first two layers with Pinwheel Convolutions (PConv). By using asymmetric padding and parallel convolution kernel design, the receptive field is effectively expanded, the ability to capture underlying features is enhanced, and the detection of small diseased areas in eggplants is more precise. In the feature fusion stage, this paper designs a Cross-layer Attention Module (CAM), including Cross-layer Channel Attention (CCA) and Cross-layer Spatial Attention (CSA), which can efficiently interact and fuse features of different scales without additional sampling, alleviating the information loss caused by semantic gaps. In addition, to solve the instability caused by IoU fluctuations in the bounding box regression process, the model introduces Scale-based Dynamic Loss (SD Loss), which dynamically adjusts the loss weight based on the size of the target. By adaptively adjusting the proportion of IoU-based loss and location constraint loss, more precise localization and stable regression of small diseased areas in eggplants are achieved. Experimental results demonstrate that EDDet achieves a notable improvement in mAP50 (85.4%), outperforming the baseline by 2.8%.Importantly, EDDet also Maintains excellent efficiency with only 2.75 M parameters, 9.1 GFLOPs, and a high inference speed of 288.3 FPS, which is 37.5 FPS higher than the baseline.These results highlight the model’s strong potential for real-time deployment in complex agricultural scenarios where both precision and speed are critical.