Scholarlink

<p>Quality control processes are important for farm-produced products. The need for automation in these processes is increasing, for quality grading and disease detection, which are crucial in consumable agricultural products such as potatoes. This study aims to develop a multi-task learning (MTL) deep learning model that simultaneously tackles disease detection and quantitative quality grading in potato tubers. For this purpose, a dataset of 1692 images collected from five sites and labeled by experts based on seven key attributes was compiled. YOLO11m object detection was employed to extract the area in the images followed by evaluating deep learning models (ConvNeXt/Base, ResNet-50, and EfficientNet-B2) using transfer learning. The proposed MTL model was optimized by assigning a weight of 3.5 to the quality estimation task and a weight of 1 to the disease classification task. Experimental findings revealed that the ConvNeXt/Base architecture outperforms the models. The experimental evidence indicates that the multi-task learning approach can be deployed as a reliable tool for automating potato quality assessment, delivering high predictive accuracy while reducing output variability compared with separate single-task models.</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

Potato Quality Grading and Disease Testing with Artificial Intelligence–Assisted Image Processing Methods

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Pınar Koç,
  • Caner Balım

摘要

Quality control processes are important for farm-produced products. The need for automation in these processes is increasing, for quality grading and disease detection, which are crucial in consumable agricultural products such as potatoes. This study aims to develop a multi-task learning (MTL) deep learning model that simultaneously tackles disease detection and quantitative quality grading in potato tubers. For this purpose, a dataset of 1692 images collected from five sites and labeled by experts based on seven key attributes was compiled. YOLO11m object detection was employed to extract the area in the images followed by evaluating deep learning models (ConvNeXt/Base, ResNet-50, and EfficientNet-B2) using transfer learning. The proposed MTL model was optimized by assigning a weight of 3.5 to the quality estimation task and a weight of 1 to the disease classification task. Experimental findings revealed that the ConvNeXt/Base architecture outperforms the models. The experimental evidence indicates that the multi-task learning approach can be deployed as a reliable tool for automating potato quality assessment, delivering high predictive accuracy while reducing output variability compared with separate single-task models.