Abstract: Aiming at the problem that the age estimation accuracy of face images in terms of spatial and time series fine-grained features is not high, this paper proposes a convolution long-short term memory (ConvLSTM) network model based on the attention mechanism. We introduced an attention mechanism between the upper hidden state of the ConvLSTM model and the current input state to increase the weight of feature factors that had a significant impact on age estimation. The channel weight factor was obtained by averaging pooling, and the attention weight was normalized to obtain a new input state. The feature extraction and age estimation were realized by the new input state and ConvLSTM model. In order to verify the effectiveness of the model, FG-NET and MORPH face datasets were used as experimental objects, with mean absolute error (MAE) and cumulative score (CS) as evaluation indicators. The experimental results show that the average absolute errors of the algorithm model on the FG-NET and MORPH face data sets are 3.60 and 2.45, respectively; the cumulative index reached 89.3% on the MORPH data set. Compared with the non-attention ConvLSTM model and the LSTM model, the cumulative index is increased by 0.80 and 4.60 percentage points. It also has a good performance in the complexity of the algorithm model.