Building Extraction And Deep Learning In Remote Sensing Images

With the fast development of technology,an increasing number of images are sent back to earth timely every day,which 70 percent images are about urban cities.Among them,the artificial building is about 70%.Although the computer and science have made great process,fast and robust building extraction from high-resolution image is still an activity field in relevant researches.Based on artificial designed features,the original building extraction algorithm has a better method on building extracting with specific shapes,but it needs to improve processing effects in complex situations.Recent development of deep learning has solved the problem of object recognition and detection.Therefore,a novel extraction method for building extraction by using deep learning was proposed.I adopt the idea of first classifying then detecting,and use convolutional neural networks to construct buildings.Through the self-built data set training,this method has solved the problem of building extraction in high-resolution images.The paper mainly focuses on using deep learning method to detect building from high-resolution remote sensing images.The images used were from Google Earth and Baidu map as experimental data.First,two networks are designed and trained to test the depth of network on accuracy,and then changing one of the networks with some small improvements to test how each one of the layers influence the accuracy,and based on the situation how to improve the accuracy.Then the author finds that in the case of insufficient data,using the pre-training model to fine-tune the dataset can achieve a higher accuracy compared with the networks designed on my own.After that,using deconvolution to check what features the network learned and how to adjust the hyper-parameters as a reference.After that,the author uses a convolutional neural network called Network 2 as the feature extractor and improves the existing Fast R-CNN algorithm and uses it in building detection.The author trains the two detection networks at the same time.And the results of the two networks are fused,it can detect the buildings the image quickly and accurately.On contrast with the Fast R-CNN method,the method in the paper achieves a higher accuracy and error-tolerant ability with the condition of no speed lose.The algorithm processes of this paper are: data preprocessing,building candidate area extraction,candidate area identification,candidate area building boundary regression.Finally,I use C++ and python to complete the data annotation,data serialization and network training reasoning experiments.This experiment proves that the high-resolution remote sensing image building extraction algorithm can complete the detection and identification quickly and accurately with deep learning.Compared with the Fast R-CNN algorithm,the method proposed in the paper is much more accurate with the condition of no speed reduction.Besides,the robust and error-tolerant ability of the network proposed in the paper is much higher,and almost meet the requirements of real-time detection.Compared to the existing research results,both extraction accuracy and the algorithm efficiency are improved,the application conditions are more relaxed.This also shows the effectiveness of this method in the paper.