A plug-in that simulates gaze and glance functions

The convolutional neural network gradually extracts the local features of the image through several convolutional layers. Studies have shown that the spatial correspondence of these features is particularly important. In the biology field, studies have shown that compared with human visual saliency, the Saccade path of the human's eye can better reflect human visual search behavior, and deepen the impact on dynamic visual attention. Therefore, to further improve the representation ability of the network, this paper constructs a network structure of Glance From top to bottom network (Gftb-Net), which can simulate the gaze of the human eye and stimulate the glance function and stimulate the brain's long-term and short-term memory mechanism, the network can selectively retain the characteristics of important features to further build interdependencies, adaptively change the feature information at the row and column level of the picture, and further improve the ability to collect global information. Based on the correlation between image rows and columns, a channel interaction strategy based on LSTM for collecting images to give weights between rows and columns is proposed. This method effectively combines the relationship between image rows and columns with spatial attention to extracting effective weight information. Then, the matrix dot product is combined with the attention vector after global pooling and standard deviation pooling, which effectively improves the representation ability between image rows and columns. The network's top1_accuracy has been reached at 79.25% in the Imagenet-100 classification task. Experiments have proven that the network is effective, and the plug-in can theoretically be used after any convolutional layer.