House Recognition Based On GF-2 Remote Sensing Image Of Deep Learning

Using remote sensing image to identify houses is a hot spot of remote sensing image recognition.With the implementation of high score project in China,high score data source has become an important data source of remote sensing data in China.GF-2 satellite is the first high-resolution satellite in China.With high spatial resolution,GF-2 remote sensing image has been widely used in ground object recognition.How to effectively identify houses in GF-2 image and obtain information in GF-2 image is of practical significance.In recent years,the deep learning algorithm has achieved great success in the field of video and image recognition.In this paper,we use the deep learning model to identify houses in the GF-2 remote sensing image and count the number of houses.Firstly,the houses in a GF-2 were selected as samples,and after the samples were processed,the house data set of the GF-2 was established.Secondly,ZF and VGG16 networks were used in Faster RCNN model to identify houses in the GF-2 and compared with traditional methods.After that,the house data set of GF-2 was improved,and the accuracy of house identification using Faster RCNN and Mask RCNN model was compared.Finally,the multi-scale candidate frame was used to optimize the region candidate network in the Mask RCNN model,and the precision and recall ratio of the optimized Mask RCNN model were discussed.Main contents of this paper are as follows:According to the different characteristics of different feature extraction networks in the same depth learning model,this paper selects the feature extraction network which is suitable for the recognition of GF-2 houses under the fast RCNN model.Based on the data set of GF-2 houses,the network training was carried out.The feature extraction network suitable for GF-2 house recognition was selected from ZF and VGG16 networks,and compared the results with the traditional method.The experimental results show that the accuracy of Faster RCNN(VGG16)model is 2.75% higher than that of Faster RCNN(ZF)model,and the recognition effect is better than that of the traditional method.In view of the problem that the original GF-2 image house data set can not reflect the actual imaging situation of the housing,this paper proposeed an improved method of GF-2 data set based on noise.Based on the house data set of GF-2,the network training was carried out.The feature extraction network suitable for GF-2 house recognition was selected from ZF and VGG16 networks,and compared with the traditional method.The experimental results show that the improved data set improves the accuracy of house identification by the Mask RCNN model and the Faster RCNN(VGG16)model,and the accuracy of house identification by the Mask RCNN model is higher.Aiming at the problem that the scale of candidate frame is sole in the area candidate network of Mask RCNN model,and the house feature extraction ability of in GF-2 is weak,this paper proposeed a multi-scale area candidate network optimization method for house extraction in GF-2 image based on Mask RCNN model.The multi-scale candidate frame was used to replace the single scale candidate frame in the original network to optimize the regional candidate network in the Mask RCNN model.This paper discussed the accuracy of the optimized Mask RCNN model for house identification,and compared it with the Faster RCNN(VGG16)model and the Mask RCNN model.The accuracy and recall ratio of the optimized Mask RCNN model for house identification are 92.73% and 89.14% respectively.The experimental results show that the optimized Mask RCNN model is more suitable for house identification in GF-2 image.