| With the rapid development of artificial intelligence and computer vision,data-driven methods based on deep learning have achieved significant applications in practical problems.Some studies suggest that there is no general AI network.Therefore,how to extract features with representational ability from data combined with domain knowledge for a specific problem becomes the key to solving the problem.In this paper,aiming at the two tasks of tropical cyclone intensity estimation and scene structural reconstruction,combined with the specific domain knowledge in the two tasks,a unique feature extraction and learning method based on attention mechanism is designed,and the features with representational ability are extracted and used in the experimental results have been confirmedIn the problem of tropical cyclone estimation,the DR-transformer model proposed in this paper can extract features of distance consistency and rotation invariance from tropical cyclone images,and can better learn the contour,structure and other visual features of each tropical cyclone image,which leads to smaller estimation error between adjacent intensities.We further use the transformer module to learn the temporal correlation between its features.The experimental results show that the framework proposed in the paper achieves the highest results on public datasets.In the problem of scene structured reconstruction,the paper creatively uses line cloud data as input and constructs a related line cloud dataset.After that,our proposed framework takes the line patch which containing enough local spatial information as the input.The paper obtains the line patch feature by learning the spatial correlation between the line segments in the line patch,and finally uses this feature to perform the tasks of vertex prediction and connection relationship prediction.The experimental results show that the performance on the test set proposed in the paper is better than other baseline experiments. |
PDF Full Text Request