The imaging results of positron emission tomography (Positron Emission Tomography, PET) equipment are often constrained by some factors such as tracer dose and scanning time, resulting in the image quality decline and affecting doctors’ diagnostic results. At present, improving the quality of PET imaging with artificial intelligence (AI) technology is a hot research topic. This paper aims at the problems of existing methods such as many training parameters, loss of shallow information, loss of texture details, etc., and proposes a method based on residual hybrid domain attention. The PET super-resolution reconstruction method of force network. This method designs a lightweight convolutional network, in which the residual learning structure is added and the mixed domain attention block is incorporated. While enhancing the interaction of the neural network, it also increases the attention to the high-frequency information area then quickly reconstruct the high-frequency details of the image. The data set includes open source data in the network and clinical data obtained from hospitals. As a result, a super-resolution data set of PET images is established for training and testing. The experimental results show that the test results of the algorithm in this paper and the comparison network are significantly improved. When the scale factor is 4, compared with CARN (Cascading residual network), the average values ​​of PSNR and SSIM are increased by 0.09dB and 0.0009, respectively. In addition, the number of parameters is reduced by 50.26%, which effectively improves the reconstruction efficiency of the model.