Scholarlink

<p>Thermal tomographic image reconstruction has emerged as a crucial technique for non-destructive assessment, predictive maintenance, and quality control across a wide array of industrial applications. However, classifying such images remains a complex task, primarily due to the intricate nature of temperature distribution patterns and the limited availability of annotated datasets. This study presents a systematic investigation into the efficacy of deep transfer learning approaches for the classification of thermal tomographic images, with a specific focus on the estimation and evaluation of heating non-uniformity within an enclosed contour of a hot airflow chamber. Thermal tomographic image datasets are synthetically generated using COMSOL Multiphysics, employing the ‘finite element method’ (FEM) to simulate various heating and fluid-flow patterns based on boundary flex and thermistor sensors measurements. A comparative evaluation has been conducted using seven widely adopted pre-trained ‘convolutional neural network’ (CNN) architectures such as AlexNet, ResNet18, ResNet50, VGG19, SqueezeNet, DenseNet201, and InceptionV3. These models have been employed to extricate discriminative features and perform the classification of thermal patterns indicative of temperature distribution anomalies. Accuracy, precision, recall, and F1-score are among the common classification metrics used to evaluate each model's performance. Among the evaluated architectures, SqueezeNet demonstrated superior performance, attaining the highest accuracy of 92.81% and an F1-score of 92.83%, whereas AlexNet yielded the lowest accuracy of 89.38% with F1-score of 89.38%. Moreover, to achieve reliable and robust performance out of these models, a five-fold cross-validation testing technique has been employed. The results affirm the applicability of deep transfer learning for enhancing the interpretability and classification accuracy of thermal tomographic images. The proposed framework offers a generalized and robust methodology for assessing heating non-uniformity, with potential deployment in automated inspection systems for thermally-driven fluid-flow processes.</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

Deep transfer learning framework for thermal tomographic image classification: a comparative study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月16日

查看全文
  • Chayan Kumar Basak,
  • Gautam Sarkar,
  • Palash Kumar Kundu

摘要

Thermal tomographic image reconstruction has emerged as a crucial technique for non-destructive assessment, predictive maintenance, and quality control across a wide array of industrial applications. However, classifying such images remains a complex task, primarily due to the intricate nature of temperature distribution patterns and the limited availability of annotated datasets. This study presents a systematic investigation into the efficacy of deep transfer learning approaches for the classification of thermal tomographic images, with a specific focus on the estimation and evaluation of heating non-uniformity within an enclosed contour of a hot airflow chamber. Thermal tomographic image datasets are synthetically generated using COMSOL Multiphysics, employing the ‘finite element method’ (FEM) to simulate various heating and fluid-flow patterns based on boundary flex and thermistor sensors measurements. A comparative evaluation has been conducted using seven widely adopted pre-trained ‘convolutional neural network’ (CNN) architectures such as AlexNet, ResNet18, ResNet50, VGG19, SqueezeNet, DenseNet201, and InceptionV3. These models have been employed to extricate discriminative features and perform the classification of thermal patterns indicative of temperature distribution anomalies. Accuracy, precision, recall, and F1-score are among the common classification metrics used to evaluate each model's performance. Among the evaluated architectures, SqueezeNet demonstrated superior performance, attaining the highest accuracy of 92.81% and an F1-score of 92.83%, whereas AlexNet yielded the lowest accuracy of 89.38% with F1-score of 89.38%. Moreover, to achieve reliable and robust performance out of these models, a five-fold cross-validation testing technique has been employed. The results affirm the applicability of deep transfer learning for enhancing the interpretability and classification accuracy of thermal tomographic images. The proposed framework offers a generalized and robust methodology for assessing heating non-uniformity, with potential deployment in automated inspection systems for thermally-driven fluid-flow processes.