Impact of Lighting-Based Biases on the Performance of YOLOv8 Object Detection Models

Object detection systems play a crucial role in numerous applications, ranging from autonomous vehicles to medical imaging. These systems must perform reliably under a variety of environmental conditions, including variable lighting, which can affect detection accuracy. This study investigates the influence of lighting-based biases on the performance of the YOLOv8 object detection model. A controlled image generation process was used to isolate the effects of brightness on detection accuracy by transposing objects onto randomly generated backgrounds and systematically varying brightness levels. A comprehensive dataset comprising 192,000 images across 40 brightness levels was created to assess the impact of lighting variations. Results demonstrate that model accuracy is improved at the higher brightness levels. Surprisingly, accuracies at the lowest and highest brightness levels were much higher than intermediate brightness levels, with a high statistical significance (p < 0.0001). Our findings emphasize the importance of addressing lighting-induced biases to improve the reliability of AI-driven object detection systems. Future work can focus on examining other environmental factors such as glare, contrast, and color variations, as well as validating these methods in uncontrolled environments.