Automatic Recognition Of Buildings In UAV Remote Sensing Images Based On Improved Mask-RCNN Algorithm

In the era of rapid development of artificial intelligence,image recognition plays an important role in various fields.Automatic recognition of buildings in remote sensing images is a research hotspot in the fields of remote sensing surveying and computer vision.The application of deep learning to automatic recognition of buildings in remote sensing images can reduce the work of artificial vectorization,provide basic data for residential construction,statistics and other fields,and play an important role in urban spatial planning and land use efficiency.In this paper,through the further research and analysis of the traditional Mask-RCNN image recognition algorithm,focusing on the low efficiency and low accuracy of label labeling and the accuracy,accuracy and recall rate in the recognition process of the traditional Mask-RCNN algorithm in the remote sensing image building recognition,from the label labeling and feature image extraction two aspects,the live wire algorithm and the traditional Mask-RCNN image recognition algorithm is improved to improve the performance and optimize the recognition effect in the process of remote sensing image building recognition.Finally,the efficiency of the algorithm is verified by experiments.The specific work is as follows:(1)Aiming at a large number of tags needed in deep learning,in order to overcome the difficulties of UAV platform's complex motion,insufficient lighting conditions and complex surface feature contour,an improved live wire algorithm is proposed and applied to label typical surface features of UAV remote sensing images.This method combines the improved Pal-King's fuzzy edge detection method and optimizes the cost function by increasing the change feature of gradient amplitude between nodes.It achieves the robustness and higher efficiency of tag extraction,and the effectiveness of this method is verified by experiments.(2)In view of the loss and confusion of image features in the process of remote sensing building recognition by traditional Mask-RCNN algorithm,combined with the special features of remote sensing buildings,an improved Mask-RCNN remote sensing building recognition algorithm is proposed.In this method,a feature fusion layer is added to the FPN layer to fuse the strongest position information layer and the strongest semantic information layer.By combining the down sampling with the existing up sampling,the possible impact of feature loss is reduced,and the receptive field is expanded by using the 3 * 3 hole convolution kernel to reduce the confusion effect.Finally,combined with the proportion characteristics of remote sensing buildings,the proportion of regional prediction box is adjusted appropriately to achieve the high efficiency of remote sensing image building recognition results.The effectiveness of the improved algorithm is verified by experiments.