An optimized YOLO NAS based framework for realtime object detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Object detection models are crucial for various AI applications.
• Existing YOLO-NAS variants require further optimization for enhanced performance.
• Integrating advanced activation functions and optimization algorithms can improve model stability and accuracy.

Purpose of the Study:

• To enhance the YOLO-NAS object detection model by incorporating MISH activation and Artificial Bee Colony (ABC) optimization.
• To evaluate the performance of the enhanced YOLO-NAS model against baseline YOLO-NAS variants and other state-of-the-art models.
• To demonstrate the effectiveness of combining biologically inspired optimization with advanced activation functions for improved object detection.

Main Methods:

• Integration of MISH activation function for improved feature representation and gradient flow.
• Application of Artificial Bee Colony (ABC) optimization algorithm for hyperparameter tuning.
• Testing the enhanced YOLO-NAS model on a custom dataset and comparing its performance metrics (precision, recall, mAP) against YOLOv6, YOLOv7, and YOLOv8.

Main Results:

• The enhanced YOLO-NAS model demonstrated superior performance across precision, recall, and mean average precision (mAP) compared to baseline YOLO-NAS variants.
• The proposed model outperformed YOLOv6, YOLOv7, and YOLOv8 in accuracy, recall, precision, F1 score, and mAP at various intersection over union (IoU) thresholds (0.50, 0.75, 0.95).
• The fine-tuned model achieved a remarkable 98% accuracy in real-time object recognition tasks.

Conclusions:

• The combination of MISH activation and ABC optimization significantly enhances YOLO-NAS model training stability and prediction accuracy.
• The proposed fine-tuned YOLO-NAS model represents a state-of-the-art approach for real-time object detection.
• This research highlights the potential of integrating bio-inspired optimizers with modern activation functions for advancing computer vision.