Dexterous control of robotic hand driven by human motor intent has drawn a lot of attention in both industrial and rehabilitation scenarios. Providing simultaneous and proportional control has become a prevailing solution recently. Towards improving the finger kinematics estimation precision and reducing its computational cost, a convolution model with attention mechanism (CNN-Attention) was proposed in this study. For comparison purpose, two previously used deep learning models, the long short-term memory (LSTM) and the Sparse Pseudo-input Gaussian processes (SPGP) were also included. By using surface electromyography (sEMG) and kinematic signals corresponding to six hand grasp movements, the estimation performance of each of the three models was evaluated with three measures, Pearson Correlation Coefficient (CC), Root Mean Square Error (RMSE), and coefficient of determination (R2) between real and estimated joint angles. The results demonstrated that the proposed CNN-Attention model outperformed LSTM and SPGP significantly (p-value<0.05), with an average value of CC, RMSE, and R2, 0.87, 9.65 degrees, and 0.73, respectively. Also, the CNN-Attention model is more stable and versatile over various subjects and joint angles in comparison to other models. Additionally, the computational time to build a CNN-Attention was obviously shorter than that to train a LSTM model (43.00 ± 4.25 min vs. 73.40 ± 5.81 min). These findings suggest that the CNN-Attention would be a promising model for continuous estimation of hand movements in the human-machine interaction and cooperation.