Surface Features Extraction On Mining Area Image Based On Object-oriented And Deep-learning Method

Coal resource plays an important role in promoting economic development,and how to ensure the safety of production in mining areas is the top priority in mining operation.Accurate and efficient acquisition of surface features on mining areas can effectively assist the arrangement and deployment of the work in mining areas.However,surface features in mining areas are always special and complex,so it takes time and effort to obtain information of large mining area by traditional manual investigation methods.In order to obtain the surface feature information of coal mining area quickly and accurately,this paper used UAV flying at low altitude to obtain high definition image data in the mining area and proposed an object-oriented combined with deep-learning classification method to extract surface features.In order to make labels for deep learning semantic segmentation,this paper proposed a labeling method using object-oriented classification with manual correction,which improved the accuracy and speed of labeling.As for the deep learning method,this paper implemented 4 convolutional neural network models(i.e.VGG-16,VGG-19,ResNet-34 and Xception)and 3 full convolutional neural network models(i.e.FCN-32s,FCN-8s and U-Net).Experiments showed that different models had different extraction effects on specific types of surface features,to improve classification accuracy,2 integrated algorithms,which were majority voting algorithm and scoring algorithm based on these deep learning models were proposed.At last,the edge of the surface features was optimized by the dilation and erosion algorithm.The experimental results showed that,the overall accuracy of feature extraction methods based on convolutional neural network was around 80%,which was much worse than that based on the full convolutional neural network,and this proved that the convolutional neural network models were not suitable for extracting such small size surface features Compared with the single object-oriented classification method and the convolutional neural network method,the method based on full convolutional neural network and object-oriented classification had higher surface feature extraction accuracy and higher Kappa coefficient.Besides,the integrated algorithm took advantage of different models and combined the characteristics of different models which had good effects on specific surface feature extraction,so as to improve the surface feature extraction accuracy.Finally,the dilation and erosion algorithm effectively eliminated the noise pixels of the surface feature extraction results and optimized the extraction effect.Among all methods,the scoring integrate model based on the full convolutional neural network models had the best recognition effect.The overall accuracy of feature extraction on the testing image data set was 94.55%,the Kappa coefficient was 0.8191.