In recent years, deep learning methods have made significant progress in the field of speech enhancement, but the existing models generally have the problems of redundant parameters and high inference overhead, which are difficult to meet the requirements of edge computing devices for low resource occupation and real-time performance. In order to solve this problem, an efficient and lightweight speech enhancement framework was proposed. The front-end of the model uses the band merging to compress high-frequency redundant information. Group fusion convolution was introduced, and the dilated deep convolution and time-frequency hybrid dilated residual attention module were combined to improve the multi-scale modeling ability. At the same time, the cross-channel feature interaction is enhanced by the channel aggregation module. The group dual-path long short-term memory module is used in the enhancement layer, which combines channel grouping and dual-path mechanism, which not only maintains excellent time-frequency modeling capabilities, but also realizes the lightweight of the model structure, and effectively reduces the number of parameters and inference overhead. Experiments are carried out on the VoiceBank DEMAND dataset, and the results show that the voice perception quality of the proposed model is 2.99, the parameter quantity is 51.85 K, and the computational amount is 145.23 MMac, which achieves a good balance between performance and computing efficiency.