Research On Matching Area Selection Of Remote Sensing Image Based On Convolutional Neural Networks

The technology of automatic selecting a matching area plays a significant role in aircraft navigation and ground information assurance, which directly affects the quality and accuracy of matching. How to make use of satellite images has been an emphasis as well as a difficulty in the field of accurately matching guidance. However, compared to the normal object detection, it is quite difficult to describe and quantify the characteristics of a matching area due to its various status and no stable structure. Consequently, it is extremely difficult to select a matching area, and the result often fails to be effective.Convolutional Nerual Networks(CNN) has the characteristics of self-learning the features by training and describing the structure. Within the context of the rapid development of CNN, this paper attempts to innovatively combine CNN with the technology of selecting matching area, and propose a CNN-based method of automatically selecting the matching area of remote sensing image. Algorithom used in this paper incorporates two phases: learning and selection. In the learning phase, the matching area with a large amount of information and a stable structure was selected as learning sample, based on which the matching area-based CNN classification model was designed and trained. In the selection phase, remote sensing images were cut into pieces of patches, the matching area of which was then recognized through the CNN classification model. Non-maxima suppression was used to filter out the low rate matching area, and then, a correlation peak-based uniqueness analysis(the ratio of primary and secondary peaks and the highest sharpness of peak) was used to ensure the matching area with no similarity pattern in the remote sensing image and label the selected matching area on the remote sensing image.Finally, a comparison experiment was conducted with traditional top-down rules-based selection methods, such as based layer and based SVM, in terms of the quality of selection and classification, the results of which were tested through a simulation experiment. The final results demonstrated the better selection quality of the method proposed in this study. This method was demonstrated to show strengths in improving the effectiveness of selection and the accuracy of classification and effectively selecting the matching area. The selection adaptive of this method was also demonstrated through adding noise, multi-scale and rotate experiments.