Camouflaged object detection (COD), which aims to accurately identify objects that visually blend into surroundings, has attracted increasing interest recently. Existing models usually seek a breakthrough in the RGB domain. However, it is difficult to distinguish the target objects that are visually consistent to the backgrounds in some challenging scenarios. Considering that the frequency components can more effectively capture the details and structures of the image, we rethink the COD task from the perspective of the frequency domain. To this end, we propose a frequency learning network to mine boundary and position cues for prediction. Specifically, we design the frequency feature aggregation module to merge cross-level frequency features, which are then grouped to generate details and position cues by the frequency feature learning module. Subsequently, we propose the frequency-assisted object-boundary calibration module and the dual-guidance feature reasoning module to progressively optimize the dual-guidance cues to help calibrate the camouflaged object feature for high-quality prediction. Quantitative and qualitative experimental results demonstrate that our network outperforms the state-of-the-art COD methods.