Geographical Annotation Of Image Based On Visual Spatial Congnitive Learning

With the proliferation of image capturing devices and sensors,social images are gradually being attached with a new metadata: geo-tag,which is more and more important in the Geographic Information System especially when tagging images with geographical tags.How to effectively exploit the high-level semantic feature and underlying correlation among different types of geographical features is a crucial task for geographical topics learning.Deep learning recently has demonstrated robust capabilities for image tagging.The features are learned and iterated directly from the original images so that the associations between output characteristics and the original images are more straightforward,and this ability makes deep learning method unique in geographical tagging task.With the help of human visual attention mechanism in selecting the area of interest in the images and giving a higher priority in processing,the deep learning method could also simulate that mechanism which allows the computer to find and locate the discriminative region quickly.Therefore,the visual attention mechanism is brought into the deep learning in this thesis.First,geographic features learning and abstracting with the convolution neural network is explored.Then the discriminative regions in the image are located with the global average pooling.The Spatial Activation Map(SAM)is used to realize the task of learning the geographical features through the spatial cognition in the convolution neural network.Finally,the relevant geographical tags of the images are given.Its specific work mainly includes the following contents:1.Based on the human visual cognitive mechanism,the influence of spatial cognition in the process of selecting the discriminative region is discussed.The Spatial Activation Map is generated to prove that the convolutional neural network has chosen the discriminative region with the influence of visual attention mechanism to recognize and classify images.2.With the experience in deep learning,a specific convolutional neural network is designed and the visual attention mechanism is applied to the designing process.The network automatically learns the geographical features of the images from a large amount of input data.Combined with the global average pooling,the discriminative regions in images are selected by the network,and the geographical tags are given as the output.Based on the experiments under the framework of TensorFlow,it is found that the discriminative regions in images could be effectively extracted by the convolution neural network based on spatial attention,which provides a new research idea for geographical learning,recognition and labeling.