Adaptive Region-merging Method And Application In Remote Sensing

Superpixel technology has been successfully applied in the fields of image processing,satellite remote sensing and computer vision due to its fast segmentation,clear target boundary and human vision.However,due to the initial segmentation of superpixels,the problem of "under-segmentation" and "over-segmentation" inevitably arises,and this phenomenon is particularly prominent in the case of hyper-spectral images.On the one hand,rapid advances in remote sensing technology have led to clearer targets for images features;on the other hand,the lagging development of image processing technology has made accurate target segmentation more difficult.How to make full use of the boundary consistency of superpixels to achieve ready segmentation of feature targets is becoming a popular problem in visual remote sensing.In this thesis,the following work is accomplished in the post-processing of superpixel segmentation:(1)A hierarchical processing structure of "pixel-> superpixel-> image blocking->high-resolution image" is formed.The effect of superpixel technology on the segmentation of high-resolution remote sensing images is investigated,which can preserve image details and target edges without losing pixel information,and reduce the influence of image noise and redundant information on the segmentation target.(2)The unsupervised region-merging method Max Cov(Maximum Coverage)is proposed.This method does not start from the most similar regions,but adds them to the "virtual-merging" set,and then decides whether to merge them according to the integrity(global)of the target regions(called "real-merging").Theoretically,it is shown that the region-merging problem can be converted to computing the maximum coverage of a set.In addition,two speed-up strategies are proposed to further improve the merging speed.(3)A supervised region-merging method is proposed.In view of the complexity of feature targets,for example,the problem of merging multiple feature targets at the same time,this thesis proposes a supervised region-merging method based on the migration learning method,i.e.using multi-scale depth features of feature targets and multi-level segmentation to obtain remote sensing image feature boundaries and superpixel chunks,and then combining the grid classification with the voting strategy of multi-class targets to classify superpixels in order to obtain accurate remote sensing image region-merging results.To verify the validity of methods(2)and(3),extensive experiments were conducted on the BSDS500 dataset and remote sensing dataset(GID,forest fire ground remote sensing data,UAV aerial tree canopy data,Google Earth data)respectively.