Research Article
Research on the Image Description Algorithm of Double-Layer LSTM Based on Adaptive Attention Mechanism
Algorithm 1
AAM-DLSTM algorithm training process.
| Pseudo-code implementation of AAM-DLSTM algorithm | (1) | for i:1-10000{ | (2) | for batch:1-Total number of images in the training set/batch_size{ | (3) | Read a batch_size number of images into the model; | (4) | for t:1-T{ | (5) | Extract global features and spatial local features of the image | (6) | Update the hidden layer state and memory layer state of the LSTM-1 module, and output the visual attention sentinel vector vst; | (7) | Update the hidden layer state and memory layer state of the LSTM-2 module; | (8) | Generate the predicted word st for the current time step | (9) | } | (10) | Get the text description sentence S of the current image; | (11) | Update each parameter in the AAM-DLSTM model; | (12) | } | (13) | } | (14) | All the loops are over, and the parameters of the AAM-DLSTM model are obtained |
|