Fλ/d as a predictor of image classification algorithm performance.

Area of Science:

• Computer Vision
• Machine Learning
• Sensor Technology

Background:

• Machine learning algorithms like YOLO are crucial for object classification.
• Sensor performance metrics, such as the FLD (Full Width at Half Maximum) metric, can significantly influence algorithm outcomes.
• Evaluating algorithm performance under varying sensor conditions is essential for real-world applications.

Purpose of the Study:

• To investigate the impact of sensor FLD metric variations on YOLO object classification performance.
• To establish a relationship between the FLD metric and the probability of detection for YOLOv3 and YOLOv10.
• To provide recommendations for assessing algorithm performance based on sensor FLD values.

Main Methods:

• Trained YOLO_v3 and YOLO_v10 algorithms on a standard Teledyne FLIR Systems dataset.
• Degraded image quality of a separate test dataset to simulate a range of FLD metric values (0.339 to 7.98).
• Evaluated YOLO algorithm performance on the degraded dataset to determine the probability of detection at different FLD levels.

Main Results:

• Performance of YOLO_v3 and YOLO_v10 algorithms was evaluated across various image degradation levels.
• A correlation was observed between the FLD metric and the probability of detection for both YOLO versions.
• Results demonstrate reduced detection probability with decreasing FLD values (indicating poorer image quality).

Conclusions:

• The FLD metric is a significant factor influencing the performance of YOLO object classification algorithms.
• Algorithm performance evaluation should consider the sensor's FLD metric to predict real-world effectiveness.
• Recommendations are provided for utilizing the FLD metric in algorithm performance assessments.