Evaluation and analysis of image compression effect on neural network-based heart rate classification.

Area of Science:

• Biomedical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Neural network (NN)-based systems are increasingly used for heart rate (HR) estimation from facial images.
• Image compression is crucial for efficient data transmission and storage in remote health monitoring systems.

Purpose of the Study:

• To evaluate the impact of image compression (lossless PNG, lossy JPEG) on the accuracy of NN-based HR classification.
• To compare the performance of DenseNet-121, VGG-16, and Inception V3 models under different compression rates.

Main Methods:

• Facial images were compressed using lossless (PNG) and lossy (JPEG) formats.
• The accuracy of HR classification was assessed for DenseNet-121, VGG-16, and Inception V3 models using compressed images.
• The relationship between compression rates and classification accuracy was analyzed.

Main Results:

• VGG-16 demonstrated the highest performance, achieving 97.2% accuracy in HR detection.
• Lossy image compression (JPEG) showed only a slight impact on HR classification accuracy.
• Compressed images significantly reduced bandwidth and storage requirements.

Conclusions:

• Image compression techniques, including lossy methods, are viable for NN-based remote HR classification systems.
• The proposed system offers an effective solution for low-complexity, low-bitrate HR monitoring.
• VGG-16 is a suitable model for accurate HR estimation under image compression constraints.