Scholarlink

<p>In the last few years, many deep learning techniques have been introduced for fruit classification. These techniques are not suitable for mobile devices or any lightweight device. Also, they require large storage and very expensive training operations for many training conditions. There is a&#xa0;high requirement to investigate lightweight convolutional neural network models without sacrificing the classification result. In this study, the findings from several deep learning models applied in previous research for fruit classification are compared. A&#xa0;detailed description of datasets, practical implementation, and model architecture has been discussed. Also, we study various deep-learning methods for the classification of fruits and propose a&#xa0;very lightweight model using MobilenetV3. This can help the system to focus on the most important features of the input image. We also used the Hard-Swish (H-Swish) function in place of ReLu6. Swish function improves nonlinearity, but it is not efficient for mobile-related hardware devices; therefore, in the MobileNetV3 model, we use the H‑Swish function. A&#xa0;nonlinearity appropriate for all kinds of mobile applications and improved effective network-biased architecture has been addressed. The proposed model performed effectively on the stated dataset without overfitting. According to the testing results, our model performed well on two datasets, the Fruit360 and real-world datasets, achieving 99.2% accuracy on Fruit360 and 89.3% on the real-world dataset.</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

FruitClass: A Fruit Classification System Using Modified MobilenetV3 with Hard-Swish Function

  • 期刊论文

  • 发表日期：

    2025年2月19日

查看全文
  • Manorma Chouhan,
  • Partha Sarathy Banerjee,
  • Amit Kumar

摘要

In the last few years, many deep learning techniques have been introduced for fruit classification. These techniques are not suitable for mobile devices or any lightweight device. Also, they require large storage and very expensive training operations for many training conditions. There is a high requirement to investigate lightweight convolutional neural network models without sacrificing the classification result. In this study, the findings from several deep learning models applied in previous research for fruit classification are compared. A detailed description of datasets, practical implementation, and model architecture has been discussed. Also, we study various deep-learning methods for the classification of fruits and propose a very lightweight model using MobilenetV3. This can help the system to focus on the most important features of the input image. We also used the Hard-Swish (H-Swish) function in place of ReLu6. Swish function improves nonlinearity, but it is not efficient for mobile-related hardware devices; therefore, in the MobileNetV3 model, we use the H‑Swish function. A nonlinearity appropriate for all kinds of mobile applications and improved effective network-biased architecture has been addressed. The proposed model performed effectively on the stated dataset without overfitting. According to the testing results, our model performed well on two datasets, the Fruit360 and real-world datasets, achieving 99.2% accuracy on Fruit360 and 89.3% on the real-world dataset.