<p>Object detection in low-light conditions remains challenging. An intuitive approach is to preprocess images using low-light image enhancement methods. However, as these methods are optimized for human perception rather than machine vision, most of them lead to suboptimal detection performance. To address this limitation, we propose a detection-oriented enhancement network that jointly optimizes both tasks, Low-light Mamba Frequency Network (LMF-Net). LMF-Net integrates a multi-scale convolution branch into Mamba, enabling global context modeling and local feature extraction with linear computational complexity. As low-light images contain many informative low and high-frequency components, we introduce a frequency channel attention mechanism to better utilize multi-spectral features. We adopt an end-to-end training strategy by integrating LMF-Net with a detection network, whose parameters remain frozen during training. This ensures improved detection performance in low-light scenes while maintaining robustness in normal lighting. Furthermore, this training strategy allows LMF-Net to be flexibly integrated into existing detection pipelines. Extensive experiments on multiple datasets show that LMF-Net not only outperforms existing methods but also exhibits high computational efficiency.</p>