Related Experiment Video
Updated: Jun 15, 2025
06:08
Quantitative Visualization and Detection of Skin Cancer Using Dynamic Thermal Imaging
Published on: May 5, 2011
16.8K
Lightweight convolutional neural network (CNN) model for obesity early detection using thermal images
Hendrik Leo1, Khairun Saddami2,3, Roslidar2
1Postgraduate School of Engineering, Universitas Syiah Kuala, Banda Aceh, Indonesia.
Digital Health
|August 28, 2024
Summary
A new lightweight convolutional neural network (CNN) model offers accurate early obesity detection for mobile applications. This model balances high accuracy with reduced complexity, outperforming existing MobileNetV2 and nearing DenseNet performance at a fraction of the size.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Early obesity detection is crucial for public health, particularly for mobile health applications.
- Existing convolutional neural network (CNN) models for obesity detection often prioritize accuracy over computational complexity.
- There is a need for lightweight CNN models suitable for resource-constrained environments like mobile devices.
Purpose of the Study:
- To develop a novel lightweight CNN model for accurate classification of normal and obese thermograms.
- To minimize model complexity (size and computation cost) while maintaining high classification accuracy.
- To enable early obesity detection in mobile-based applications.
Main Methods:
- Modified DenseNet201 architecture by incorporating MobileNet's depthwise and pointwise convolutions.
- Reduced the depth of the dense block network to optimize performance and reduce complexity.
- Compared the proposed model against DenseNet and MobileNetV2 using classification accuracy, model size, and computation cost.
Main Results:
- The proposed model achieved 81.54% accuracy with a compact model size of 1.44 MB.
- Its accuracy is comparable to DenseNet (83.08%) but with a significantly smaller model size (1.44 MB vs. 71.77 MB).
- The model's accuracy surpassed MobileNetV2 (79.23%) with a similar computation cost (0.69 GFLOPS vs. 0.59 GFLOPS).
Conclusions:
- The developed lightweight CNN model effectively combines DenseNet201's feature extraction capabilities with MobileNet's efficiency.
- This model presents a viable solution for early obesity detection in mobile applications due to its high accuracy and low complexity.
- The findings suggest a promising direction for developing efficient AI-powered health monitoring tools.
