High Ratio Remote Sensing Image Compression Technology Based On Sparse Representation Learning

With the increase of the resolution of remote sensing images,the high resolution remote sensing images appear to have many new characteristics,such as rich geometric,structural,texture features,as well as refined spectrum,multi-scale objects and so on.However,JPEG and JPEG2000 use DCT and DWT transform respectively.It's difficult to represent geometry and texture information of complex remote sensing image with fixed bases.However,the visual perception of the image is largely affected by the local structure information of the image,which is the main difference between image blocks.Moreover,it's seems to be difficult to dig out the laten explanatory factors of complex unstructured scene.The common application of remote sensing images is mainly in disaster response,agricultural monitoring,urban planning,environmental assessment,etc.It can be seen that the above applications.Other information is of no value for the application of remote sensing imaging.However JPEG and JPEG2000 is not sensitive to these key information and unimportant information,which reduce the encoding efficiency.At the same time,traditional compression algorithms ignore sparsity in time dimension of time sequence remote sensing image.In this paper,three kinds of high ratio remote sensing image compression methods are proposed to solve the above problems and shortcomings of traditional image compression algorithms using sparse representation learning:1.A EGE dictionary learning compression algorithm based on sample binary tree is proposed.For the representation of high resolution image,which has rich information,samples are divided by complexity,different complex samples are used to train different dictionary,to improve the ability of local information representation.In compressing,a sample is represented by the corresponding complexity dictionary.In order to represent local texture information more better,atoms are selected among all samples in global with the consideration of entropy rate of atomic selection,which not only reduces the reconstruction error,but also combines the quantization encoding with sparse representation effectively.Thus,the unbalance representation effect of different samples,and the low coding efficiency caused by sparse representation are avoided.In this method,the expression of the local information in the dictionary and the effect of sparse representation are improved.It can be seen from the experimental results that this method achieves better results in subjective visual evaluation and objective numerical indicators than traditional algorithm.2.A WTA sparse deep network compression algorithm based on region of interest is proposed.Firstly,three bands are extracted from the hyperspectral image or multispectral image with which the RGB picture is synthesized,then the new picture can be decomposed by the SLIC algorithm,the FCM algorithm is choose as the cluster algorithm.Then matching categories of clustering results to access different categories of semantic information,selecting the region of interest,which is represented with the WTA-encoder algorithm,an auto-encoder composed of sparsely and sparsely spatially condition,so that the nodes of the hidden layer are sparseness.Because of its non-linear representation and multi-layer depth structure,the network can learn the implicit explanatory information of complex unstructured scenes,and further enhance the representation of ROI region of interest.Since we do not care about non-interested areas,the compression ratio of non-interested areas can reduce storage space by sacrificing accuracy.At the same time,in order to use the superpixel segmentation classification information,we use the SOMP algorithm to represent the non-interested region,which can reduce the number of coding bits corresponding to the coding coefficient.The final experimental results show that under the same BPP,the numerical index of this method is obviously higher than that of the traditional compression method.3.A time sequence remote sensing image compression algorithm based on self-representation is proposed.The algorithm uses the sparseness of time sequence remote sensing image in the time dimension,and selects the key time sequence through the self-representation model to reconstruct the time sequence remote sensing image improving the compression efficiency.We use a tree structure dictionary to represent the key time sequence because of its great effect on reconstructing other bands.Tree structure dictionary learning is a dictionary learning method that represents errors.The way based on representation error makes the learning more targeted with reducing the representation error.The final experimental results show that under the same BPP,the numerical index of this method is obviously higher than that of the traditional compression methods.