Scholarlink

The system of insect classification is of utmost importance in fields such as agriculture, ecology, and the study of the earth’s species and their distribution. Classifying entities, objects, or documents based on one’s knowledge and perception is inefficient and also not very accurate as it is time-consuming. The latest development of deep learning, particularly the convolutional neural networks (CNNs), has significantly enhanced the aspect of image classification as an application for automated insect identification. In this paper, we focus on utilizing MobileNetV3, which is a modern and advanced CNN network, to improve insect classification with higher accuracy and less time consumption. MobileNetV3 is best suited for mobile and edge computing applications that suggest its applicability for real field conditions. Here, we describe data acquisition and processing of insect images, training, and architecture of the considered model, as well as optimization methods. This is why our approach involves transfer learning and fine-tuning of models with an aim at enhancing the models’ performances despite the limited availability of data. The performance of the model is then established by metrics common with binary classification models, such as the accuracy, precision, recall, and F1 score. The findings also show that classification performance has been enhanced by at least thirty percent, if not more, compared to conventional approaches. Furthermore, we discuss the real-world issues concerning the necessary real-time mobile device performance or the manner of making our model portable. Examples provided by cases and the use of signals show how the model is effective in different circumstances of environmental monitoring. Thus, the study shows the applicability of MobileNetV3 in the classification of insects to significantly affect a range of researchers and practitioners in the domains of agriculture and ecology. As for future work, it will continue by mitigating the current difficulties and extending the model’s effectiveness with data acquisition and computing methods improvement.

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

Enhancing Insect Classification with Convolutional Neural Networks and MobileNetV3 Architecture: A Deep Learning Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ayan Sar,
  • Ankit Vishnoi,
  • Tanupriya Choudhury,
  • Sumit Aich,
  • Subhangi Sati,
  • Purvika Joshi

摘要

The system of insect classification is of utmost importance in fields such as agriculture, ecology, and the study of the earth’s species and their distribution. Classifying entities, objects, or documents based on one’s knowledge and perception is inefficient and also not very accurate as it is time-consuming. The latest development of deep learning, particularly the convolutional neural networks (CNNs), has significantly enhanced the aspect of image classification as an application for automated insect identification. In this paper, we focus on utilizing MobileNetV3, which is a modern and advanced CNN network, to improve insect classification with higher accuracy and less time consumption. MobileNetV3 is best suited for mobile and edge computing applications that suggest its applicability for real field conditions. Here, we describe data acquisition and processing of insect images, training, and architecture of the considered model, as well as optimization methods. This is why our approach involves transfer learning and fine-tuning of models with an aim at enhancing the models’ performances despite the limited availability of data. The performance of the model is then established by metrics common with binary classification models, such as the accuracy, precision, recall, and F1 score. The findings also show that classification performance has been enhanced by at least thirty percent, if not more, compared to conventional approaches. Furthermore, we discuss the real-world issues concerning the necessary real-time mobile device performance or the manner of making our model portable. Examples provided by cases and the use of signals show how the model is effective in different circumstances of environmental monitoring. Thus, the study shows the applicability of MobileNetV3 in the classification of insects to significantly affect a range of researchers and practitioners in the domains of agriculture and ecology. As for future work, it will continue by mitigating the current difficulties and extending the model’s effectiveness with data acquisition and computing methods improvement.