Scholarlink

<p>This study aims to detect bruises on peaches using deep learning and Near-Infrared (NIR) imaging. The research presents an extensive dataset of NIR images, trains advanced Convolutional Neural Networks (CNN) models, and develops a resilient model for detecting naturally formed bruises on bicolored peaches. The dataset includes images of both healthy and bruised peaches of the “Hale” cultivar. An experimental setup was created for noise-free image acquisition on a black background, and image processing was used to crop the original image. The final dataset split into training, validation, and test images were used to train various state-ofart models. The optimal hyperparameters were selected based on their frequent usage in fruit classification problems. The results show that DenseNet121 outperformed all the other state-of-the-art models by obtaining high accuracy and was selected as a base model for further improvement. The proposed model was crafted by integrating extra layers like the average pooling layer, dropout layer, dense layer, and SoftMax layer into the DenseNet121 architecture. Moreover, to boost its performance, an adaptive learning rate employing an exponential decay schedule was implemented. The results demonstrate promising accuracy in identifying bruise areas on peach fruit, indicating potential for enhancing quality control in the peach industry.</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

A deep learning-based bruise detection model for peaches using the NIR imaging dataset

  • 期刊论文

  • 发表日期：

    2025年8月20日

查看全文
  • Zeynep Ünal

摘要

This study aims to detect bruises on peaches using deep learning and Near-Infrared (NIR) imaging. The research presents an extensive dataset of NIR images, trains advanced Convolutional Neural Networks (CNN) models, and develops a resilient model for detecting naturally formed bruises on bicolored peaches. The dataset includes images of both healthy and bruised peaches of the “Hale” cultivar. An experimental setup was created for noise-free image acquisition on a black background, and image processing was used to crop the original image. The final dataset split into training, validation, and test images were used to train various state-ofart models. The optimal hyperparameters were selected based on their frequent usage in fruit classification problems. The results show that DenseNet121 outperformed all the other state-of-the-art models by obtaining high accuracy and was selected as a base model for further improvement. The proposed model was crafted by integrating extra layers like the average pooling layer, dropout layer, dense layer, and SoftMax layer into the DenseNet121 architecture. Moreover, to boost its performance, an adaptive learning rate employing an exponential decay schedule was implemented. The results demonstrate promising accuracy in identifying bruise areas on peach fruit, indicating potential for enhancing quality control in the peach industry.