Multi-scale Remote Sensing Image Fusion Algorithm Based On Saliency Detection And Edge Decision

Remote sensing image fusion is an important branch of image fusion.With the development of modern satellite technology,more and more sensors are mounted on the satellite.In order to get the image of comprehensive information,remote sensing image fusion technology is proposed.Remote sensing image fusion technology has many directions.This paper studies the optical-spectral image fusion.Specifically,the images obtained from satellites are panchromatic images with spatial resolution and multispectral images with spectral information.In order to obtain images with rich spatial information and spectral information,optical-spectral remote sensing image fusion technology has been proposed.Fused remote sensing images are widely used in military reconnaissance,geographic exploration,urban planning and many other fields.Firstly,the concept and research status of remote sensing image fusion algorithm are introduced and analyzed,and the principle of NSCT,a multi-scale decomposition tool,is introduced and explained in detail.Secondly,a multi-scale remote sensing image fusion algorithm based on NSCT transform is proposed.For the low-frequency sub-band information decomposed by NSCT transform,this paper adopts a low-frequency sub-band information fusion rule based on sparse representation.For the super-complete dictionary used in sparse representation,the initial dictionary is constructed by discrete cosine transform(DCT),and then the input source image is divided into blocks and transformed into vectors.Then the initial DCT dictionary is trained,and finally the ideal super-complete dictionary is obtained.This can effectively improve the adaptability of the super-complete dictionary and reduce the loss of spectral information.Aiming at the high frequency sub-band information decomposed by NSCT transform,a high frequency sub-band fusion rule based on guided filtering is proposed.First,guided filtering is used to filter the high frequency sub-band information and extract the spatial details.Then,injection method is used to fuse the high frequency sub-band.This method can effectively preserve spatial details in high frequency subbands.After that,relevant experiments are carried out.The variable parameters in the fusionalgorithm are controlled by the control variable method,and the optimal parameters are selected.Then the new fusion algorithm is compared with the comparison algorithm from both subjective and objective aspects,which proves that the algorithm is superior to the comparison algorithm in both subjective and objective aspects.Finally,this paper optimizes the previous algorithm.In the low-frequency part,a low-frequency sub-band fusion rule based on saliency detection is proposed.First,the saliency matrix is obtained based on the color and spatial contrast of the image.According to the saliency matrix,which regions of the image contain rich spectral information is determined.Then,under the guidance of the saliency matrix,the fusion is carried out in a weighted way in a specific region,which optimizes the original algorithm in low-frequency.The low efficiency of frequency fusion ensures the integrity of low frequency sub-band information.In the high frequency part,a high frequency sub-band fusion rule based on edge decision is proposed.Firstly,the high frequency sub-band is used for analysis and decision-making,and the edge region and non-edge region are obtained respectively.Then different fusion methods are adopted for different regions,and the edge region is fused by substitution.In the non-edge region,a fusion method based on the comprehensive gradient of the relevant regions as the threshold is proposed.The improved high-frequency fusion rule refines and decomposes the high-frequency sub-band information,which effectively improves the loss of information in the process of high-frequency sub-band fusion.Then,the improved algorithm is compared with the previous algorithm in both subjective and objective aspects.The experimental results show that the optimized algorithm is superior to the third chapter experiment and comparison algorithm in terms of various indicators and subjective performance.In terms of operation efficiency,the average time of the algorithm is accelerated by about 9 seconds.