In low-light conditions, detectors trained on normal-light data often experience significant performance degradation. To address this issue, low-light image enhancement methods are commonly employed to improve detection performance. However, existing human vision-oriented enhancement techniques have shown limited effectiveness, while most machine vision-oriented methods rely on standard RGB image processing or RAW space conversion, often neglecting the preservation of key object features and incurring high computational costs. To overcome these limitations, we propose an efficient low-light object detection method based on Pseudo-RAW space enhancement-LightStar-Net. This method combines a Pseudo-RAW space Enhancement module (PRE) with a lightweight network, enhancing detection capabilities for machine vision in low-light environments. Using inverse mapping to convert RGB images into Pseudo-RAW feature space, the model dynamically adjusts image enhancement parameters to optimize detection performance. On benchmark datasets such as ExDark and DARK FACE, LightStar-Net achieves outstanding accuracy and inference speed. With a simple structure requiring only 3K parameters, it significantly improves detector performance in low-light environments.