Research Of Object-oriented Remote Sensing Recognition

With the development of Remote sensing technology,the resolution of Remote sensing images has been continuously improved,and its application in many fields has attracted widespread attention.Image recognition is one of the main focuses of its research.In view of the shortcomings of traditional image recognition such as long execution time and large space occupation,this paper studies object-oriented image recognition,which mainly includes two aspects:image segmentation and image recognition.In order to meet the requirements of time,space,and precision,this paper does the following.1.Parallel transformation of super pixel segmentationMost of the traditional segmentation algorithms adopt the serial implementation,which can no longer meet the high requirements of Remote sensing image processing on time and space.this paper combines the Parallel theory for Data Blocking and the OpenMP Parallel theory for simple linear iterative cluster.According to the separable characteristics of remote sensing image data,the Remote sensing images are stripped and numbered,and the SLIC segmentation for the sub-images are independently performed.after the implementation of parallel experiments,the serial execution time and the parallel execution time are compared.Obviously,this improvement has achieved good results in both time and speedup ratios,achieving the desired goals.2.Design of edge merging algorithm After the parallel segmentation is completed,there will be a distinct merge line at the blocking edge.this paper proposes a edge merging algorithm based on the Region Adjacency Graph.this algorithm firstly obtains the region label at the boundary,establishes the region adjacency graph at the boundary according to the spatial relationship of the boundary pixels,then traverses the region adjacency graph,and processes the boundary region according to certain consolidation criteria.Finally,the effectiveness of the boundary merging algorithm is verified by the boundary merging algorithm.3.Improvement of neural network recognition algorithmDue to the effect of large Remote sensing image data and excessively large neural network training,it is difficult to achieve high precision.In this paper,Object-oriented recognition technology is used to extract spectral features and texture features of superpixels formed by image segmentation results,and combine them to form feature vectors,which need to feed into the BP neural network training.this algorithm can greatly reduce the size of the neural network.the experiment selected six remote sensing image data for vegetation recognition,compared with ENVI maximum likelihood method and Yikang KNN recognition algorithm,and the accuracy of the algorithm was verified by the accuracy.