Scholarlink

<p>Accurate and real-time quality assessment of fruits is essential for automated grading and postharvest quality control in the food industry. However, practical deployment of deep learning-based models remains challenging because surface defects are often small, localized, and visually similar to natural fruit textures, while industrial applications require both high detection accuracy and low computational cost. To address these issues, this study proposes Weighted-prior and Asymmetric-downsampling You Only Look Once (WA-YOLO), a lightweight model derived from YOLOv11n for vision-based fruit quality assessment. It first introduces a Spatial-decay-prior-weighted Convolution (wConv2d), which reweights convolutional kernels using a fixed spatial prior to enhance sensitivity to localized quality-defect cues without adding learnable parameters. Based on this operator, a Cross-stage Partial Block with a Spatial-decay Prior (C3-SDP) is designed to improve feature representation within a lightweight architecture. Additionally, an Asymmetric Downsampling (ADown) module is employed to reduce information loss during feature resolution, and Layer-adaptive Magnitude-based Pruning (LAMP) is applied to structured channel pruning to enhance computational efficiency. Experiments were conducted on a self-built fruit-quality dataset comprising 1,685 images and 2,828 annotated instances across 14 categories, including both good and defective samples of seven fruit types. This dataset was divided into 1,344 training images and 341 validation images, and model performance was evaluated using precision, recall, mAP50, mAP95, parameters, Floating-point Operations (FLOPs), model weight size, and inference speed. The results demonstrate that WA-YOLO achieves mAP50 values of 0.965 and 0.915 for good and defective fruit categories, respectively, and an overall mAP95 of 0.865, while using only 2.2&#xa0;M parameters, 3.6G FLOPs, and 4.30&#xa0;MB model weights. Moreover, the model reaches an inference speed of 166.5 Frames Per Second (FPS) under the experimental platform. Compared with YOLOv11n, WA-YOLO improves the accuracy-efficiency trade-off by reducing computational complexity while maintaining impressive detection performance. Such results indicate that WA-YOLO provides a lightweight solution for real-time assessment of fruit quality in automated grading scenarios, benefiting quality measurement and characterization in the food industry.</p> Graphical abstract <p></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

WA-YOLO: a lightweight deep learning network model for real-time measurement and characterization of fruit quality

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Jia Wen Li,
  • Wei Dong Zhang,
  • Wei Bin Lin,
  • Sheng Zhao Xiao,
  • Yue Sheng Huang,
  • Ju Jian Lv,
  • Kai Han Lin,
  • Xiang Lei Hu,
  • Lei Jun Wang,
  • Rong Jun Chen

摘要

Accurate and real-time quality assessment of fruits is essential for automated grading and postharvest quality control in the food industry. However, practical deployment of deep learning-based models remains challenging because surface defects are often small, localized, and visually similar to natural fruit textures, while industrial applications require both high detection accuracy and low computational cost. To address these issues, this study proposes Weighted-prior and Asymmetric-downsampling You Only Look Once (WA-YOLO), a lightweight model derived from YOLOv11n for vision-based fruit quality assessment. It first introduces a Spatial-decay-prior-weighted Convolution (wConv2d), which reweights convolutional kernels using a fixed spatial prior to enhance sensitivity to localized quality-defect cues without adding learnable parameters. Based on this operator, a Cross-stage Partial Block with a Spatial-decay Prior (C3-SDP) is designed to improve feature representation within a lightweight architecture. Additionally, an Asymmetric Downsampling (ADown) module is employed to reduce information loss during feature resolution, and Layer-adaptive Magnitude-based Pruning (LAMP) is applied to structured channel pruning to enhance computational efficiency. Experiments were conducted on a self-built fruit-quality dataset comprising 1,685 images and 2,828 annotated instances across 14 categories, including both good and defective samples of seven fruit types. This dataset was divided into 1,344 training images and 341 validation images, and model performance was evaluated using precision, recall, mAP50, mAP95, parameters, Floating-point Operations (FLOPs), model weight size, and inference speed. The results demonstrate that WA-YOLO achieves mAP50 values of 0.965 and 0.915 for good and defective fruit categories, respectively, and an overall mAP95 of 0.865, while using only 2.2 M parameters, 3.6G FLOPs, and 4.30 MB model weights. Moreover, the model reaches an inference speed of 166.5 Frames Per Second (FPS) under the experimental platform. Compared with YOLOv11n, WA-YOLO improves the accuracy-efficiency trade-off by reducing computational complexity while maintaining impressive detection performance. Such results indicate that WA-YOLO provides a lightweight solution for real-time assessment of fruit quality in automated grading scenarios, benefiting quality measurement and characterization in the food industry.

Graphical abstract