Hybrid pre trained model based feature extraction for enhanced indoor scene classification in federated learning...

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Indoor scene classification is vital for smart applications but challenged by complex environmental factors.
• Traditional methods like SVM and KNN offer limited performance in diverse indoor settings.
• Deep Learning (DL) models, particularly CNNs, have advanced feature extraction for improved accuracy.

Purpose of the Study:

• To propose and implement a novel MultiData model for enhanced indoor scene classification.
• To integrate DL with Linear Discriminant Analysis (LDA) and Federated Learning (FL) for superior performance and data privacy.
• To evaluate the MultiData model's effectiveness against established DL architectures and FL-based training.

Main Methods:

• Development of the novel MultiData model combining DL, LDA, and FL.
• Comparative analysis of MultiData against VGG16, VGG19, and ResNet152.
• Federated learning implementation across four clients with IID and non-IID datasets.

Main Results:

• MultiData achieved near-perfect accuracy (99.99%) and minimal validation loss (0%) compared to other models.
• Federated training demonstrated model robustness with 100% training accuracy and over 95% validation accuracy.
• The model proved effective in both IID and non-IID data scenarios.

Conclusions:

• The MultiData model offers a significant advancement in privacy-preserving indoor scene classification.
• This research supports the development of smart, sustainable environments and IoT-based automation.
• The findings are applicable to healthcare, smart infrastructure, and surveillance sectors, aligning with SDGs 9, 11, and 12.