A Multifeature Complementary Attention Mechanism for Image Topic Representation in Social Networks
Table 1
The steps of CATR: a multifeature complementary attention mechanism method.
Input: original microblog image, feature dimension
Output: fused image features
(1) The image information obtained by preprocessing unifies the resolution and pixels of the image
(2) Get the object feature of the image
(3) The focused feature of the image is extracted by VGG-19
(4) Cascade object feature Qi and focused feature of output
(5) According to equations (1) and (2), the attention distribution of the image for different objects is calculated
(6) According to equation (3), the image features under the guidance of the object are calculated
(7) Feature extraction of unfocused image based on DeepFixNet
(8) Using the same method as calculating the attention distribution of focused features, the attention distribution of unfocused features is calculated
(9) According to equation (4), the focused feature and unfocused feature are fused
(10) Use equation (5) to update the weights and parameters in the network
