Prediction of spatiotemporal sequence based on IM-LSTM
Guixin Liu, Zhonghua Ma · 2022 2nd International Conference on Computer Science, Electronic Information Engineering and Intelligent Control Technology (CEI) · 2022
With the development of deep learning, spatiotemporal prediction learning has many applications, such as weather prediction and human posture prediction. The continuous data question is described as a spatiotemporal sequence prediction question, in which both the input and the prediction target are spatiotemporal sequences. But the prediction of continuous frames increases the ambiguity of the image. A number of factors are known to affect the sharpness of image. Interaction between states in the network has a significant impact on sharpness of image. Based on the above discussion, the SIM (Short time interactive memory) and the LIM (Long term interactive memory) are designed. The current input and hidden states are updated by the SIM block. Memory cells and hidden states are updated by the LIM block. We built the I-LSTM unit by adding SIM and LIM to the LSTM. By stacking I-LSTM, we propose IM-LSTM (Interact memory LSTM). Experiments show the effectiveness and flexibility of our proposed method. In our experiments, we demonstrate that our proposed model accurately predicts future frames as well as other representations. It is superior to other advanced methods in index LPIPS.