The adaptive optical system without wavefront detection has the advantages of simple structure and easy application, and it is now turning into a research hotspot in the field of optical communication. With the rapid development of artificial intelligence technology in recent years, deep learning has been introduced into wavefront detection-free adaptive optics systems to correct wavefront aberrations. This paper proposes an adaptive optical wavefront recovery method based on the residual attention network in order to prevent the degradation of neural network. To prevent the degradation phenomenon of neural network, the residual network is first used as the backbone network, and its hopping layer connection property is utilised to enable the network model to learn deeper features. The input light intensity map is transformed into a feature map by a 7×7 downsampling convolution operation in the residual network, followed by a maximum pooling operation with a filter size of 3×3 to reduce the computational parameters and prevent overfitting phenomenon. Then, to increase the feature extraction capability of the network without significantly increasing the computational effort, this paper constructs a multi-scale residual hybrid attention network structure based on the residual network, using a null convolution operation to convert the light intensity image into a feature map for backward propagation. In the attention layer, features are extracted by convolution kernels of different scales in a distributed manner, and the dual-stream network structures of 3×3 and 5×5 sizes are chosen to extract the feature maps. The attention mechanism is used to improve the recognition rate of the network for broken light spot features and to achieve the effect of enhancing the network to express light intensity image features. Each hybrid attention module contains two convolution operations and one hybrid attention computation operation. The dimensionality of the feature map remains unchanged after the attention layer, and each channel is assigned a different weight coefficient. Finally, a network loss function combining the realistic evaluation metrics of wave crest and trough values as well as root mean square values of wavefront is designed, and the Zernike coefficients of the actual wavefront aberration are ensured in the training to match the final result. In order to simulate the transmission of vortex beams at different turbulence intensities, the Zernike coefficients and corresponding light intensity maps are randomly generated at different turbulence intensities in accordance with Kolmogorov turbulence theory. Parameters such as the gradient descent algorithm, batch size and number of iterations of the network are set reasonably, and the simulations are carried out using the keras deep learning library. The final results show that the residual attention network can reconstruct the turbulent phase quickly and accurately, and the recovered residual aberrations have peaks and troughs between 0.05 and 0.3 rad and root mean square between 0.01 and 0.07 rad. The experimental results show that the Zernike coefficients predicted by the residual attention model are similar to the actual coefficients and the phases reconstructed by the coefficients are highly similar to the actual phases compared with other network models. The effectiveness of the hybrid attention network in the task of reconstructing wavefront phases is then also verified, with the highest accuracy achieved with less increase in time complexity. The high accuracy, real-time performance and flexibility of the residual attention network provide practical applications for deep learning in adaptive optical systems.