Sequential recommendation (SR) aims to model user preferences through their historical interaction data. In recent years, Transformer has been widely used for SR due to its strong capability of modeling long-term dependencies. However, recent studies have shown that current Transformer-based SR models are susceptible to the over-parameterisation and suffer from inaccurate allocation of attention weights. Moreover, most methods model user preferences only in the time domain and cannot filter noisy input, leading to inaccurate recommendation. In this work, we propose a novel Frequency Feature Enhanced Mix Calibration Attention Network for Sequential Recommendation, named FMCARec. Specifically, we design a Squeeze-and-Excitation filtering strategy to capture the frequency features and filter noisy terms in user interaction data. Convolution operation is then used to mitigate the effects of over-parameterisation on self-attention mechanism. And a spatial information calibration strategy is used to calibrate attention weights and decoupled location coding weights. Finally, contrastive learning is utilized to improve the quality of user embedding. Experimental evaluations conducted on three public datasets demonstrate that our proposed model performs significantly better than the state-of-the-art methods.