Scholarlink

<p>To address the challenge of real-time cherry ripeness and surface defect detection, this paper proposes SEDDP-YOLO, a lightweight and optimized version of the previously proposed Cherry-YOLO model, by introducing a structure-aware distillation-guided two-stage evolutionary pruning strategy. Traditional cherry detection methods often struggle with both accuracy and deployment efficiency—particularly on resource-constrained devices—making real-time detection difficult. To overcome these limitations, we optimize the Cherry-YOLO model using a structure-aware distillation-guided two-stage evolutionary pruning strategy, reducing the number of parameters from 12.93&#xa0;M to 4.87&#xa0;M and lowering computational load from 9.1 GFLOPs to 3.7 GFLOPs. As a result, inference speed increases to 177 FPS, representing a 105.8% improvement over the original model. Experimental results show that SEDDP-YOLO achieves an mAP@0.5 of 0.909 and an mAP@0.5:0.95 of 0.635, demonstrating strong overall performance. The model offers an efficient, lightweight, and easily deployable solution for cherry ripeness and surface defect detection, with broad potential for edge computing and other resource-constrained applications.</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

SEDDP-YOLO: a dual-stage pruning and distillation approach for fast and accurate cherry quality detection

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Fei Luan,
  • Kailong Fan,
  • Xinghang Xu,
  • Xueqin Yang,
  • Jinnan Chen

摘要

To address the challenge of real-time cherry ripeness and surface defect detection, this paper proposes SEDDP-YOLO, a lightweight and optimized version of the previously proposed Cherry-YOLO model, by introducing a structure-aware distillation-guided two-stage evolutionary pruning strategy. Traditional cherry detection methods often struggle with both accuracy and deployment efficiency—particularly on resource-constrained devices—making real-time detection difficult. To overcome these limitations, we optimize the Cherry-YOLO model using a structure-aware distillation-guided two-stage evolutionary pruning strategy, reducing the number of parameters from 12.93 M to 4.87 M and lowering computational load from 9.1 GFLOPs to 3.7 GFLOPs. As a result, inference speed increases to 177 FPS, representing a 105.8% improvement over the original model. Experimental results show that SEDDP-YOLO achieves an mAP@0.5 of 0.909 and an mAP@0.5:0.95 of 0.635, demonstrating strong overall performance. The model offers an efficient, lightweight, and easily deployable solution for cherry ripeness and surface defect detection, with broad potential for edge computing and other resource-constrained applications.