Analysis of the Effect of RetinexNet-Based Image Preprocessing on Object Detection Performance Using YOLOv8 Under Low-Light Conditions

Abstract

The decline in performance in object detection systems based on deep learning is skewed to be meaningfully inclined when the system is used in low-lighting conditions, especially due to the decrease in visual quality of the image. In this study, the focus is directed to evaluate the effect of the application of RetinexNet-based image preprocessing on object detection performance using YOLOv8 in low-light environments. The experimental process was carried out to compare the detection results between models that used preprocessing and those that did not use preprocessing, based on evaluation metrics such as precision, recall, and mean average precision (mAP). The results indicate that improving the visual quality of the sword image is always followed by an increase in detection accuracy, because these changes can cause a shift in the distribution of visual features that have an impact on the model's generalization ability. In addition, the phenomenon of domain shift resulting from image changes using RetinexNet was also found, which had an effect on the consistency of YOLOv8 performance. The main contribution of this study is to provide empirical evidence that preprocessing strategies for low-light conditions not only need to focus on improving visual quality but also need to be adapted to the characteristics of the detection model in order to obtain a more adaptive pipeline under extreme lighting conditions.