A two-stage audio-visual speech enhancement algorithm based on convolution and gated attention was proposed in order to solve the problem of high complexity and poor performance of audio-visual speech enhancement model. The block-mixed gated attention unit (GAU) was utilized as the backbone of feature fusion network by using a simplified single head attention mechanism and applying intra block quadratic-inter block linear attention in order to reduce complexity and capture global dependencies of audio-video sequences. Convolutional module was integrated into GAU. Depthwise convolution and pointwise convolution were used to extract local features within and between audio-video blocks in order to capture local dependencies of audio-video sequences. The combination of convolution and attention can significantly improve the performance of the model in processing audio-video sequences. A two-stage algorithm was adopted in order to utilize the speech information contained in the two modalities. Audio was used as the dominant modality and video was used as the conditional modality in the first stage. Video in the second stage was used as the dominant modality and audio extracted in the first stage was used as the conditional modality. The experimental results show that proposed model significantly improves both PESQ and SNR metrics compared with existing models, effectively reducing complexity.