Conventional learning-based approaches for enhancing low-light images typically rely on a large amount of paired training data, which is challenging to obtain in practice. To address this issue, some algorithms have been proposed using the generative adversarial mechanism to utilize unpaired data. However, these methods commonly utilize perceptual constraints to preserve content information, leading to incorrect lightness enhancement. In this paper, we propose a dual contrastive learning scheme for unsupervised low-light image enhancement (LLIE), aiming to balance the lightness enhancement and content preservation. Specifically, we introduce two models with distinct biases towards lightness enhancement and content preservation, respectively. These models produce intermediate results that serve as negative and positive samples, guiding the final model to generate the desired outcome. Considering the coupling of luminance and noise in low-light conditions, we propose a Frequency-Spatial Attention Module to obtain an adaptive illumination map to guide light enhancement and noise removal. Extensive experimental results demonstrate our superiority over several state-of-the-art methods.