Study On Spectral-feature-based Multispectral Image Compression

Combined with spatial and spectral information, multispectral images reflect more optical characteristics of the scene than traditional chromaticity diagrams, and they have been widely used in the fields of aerospace, remote sensing, high fidelity color reproduction, environmental monitoring, art archive, biomedicine and military surveillance. As three-dimensional images, multispectral images include huge amount of data, the data storage and data transmission are significantly harder, efficient compression is required. Based on the self-features of multispectral images and for their different applications, this thesis aims to solve the problems of inadequate removal of spatial-spectral redundancy, weak adaptability, high time-space complexity, poor reconstructed colorimetric precision and so on for the existing multispectral image compression algorithms. In this dissertation, the sparse equivalent representation of multispectral images and colorimetric error criterion are proposed, efficient compression schemes by employing clustering analysis, wavelet transform, composite transformation and dictionary learning are explored, and many effective multispectral image compression algorithms, which are oriented to generic applications or color reproduction, are mainly designed.First, the theory of wavelet and wavelet transform coding methods are discussed, and based on the analysis of spatial and spectral statistical properties, the sparse equivalent representation of multispectral images is proposed, which ensures that the designed methods have better pertinency.Next, with spectral correlation of multispectral images as entry point, new algorithms named APWS and APWS_RA are designed for generic, low-complexity, and high-quality compression, which are based on spectral adaptive clustering and wavelet transform. In APWS, adaptive affinity propagation clustering, with Euclidean distance between band vectors as similarity measure, is utilized to generate inter-spectrum sparse equivalent representation which can remove inter-spectrum redundancy under low complexity. Error compensation mechanism is applied to improve the quality of reconstruction images. While, APWS_RA algorithm is formed by optimizing the rate allocation strategy of APWS. The rate pre-allocation strategy proposed in APWS_RA isdetermined by the standard deviations of sparse representation ingredients, which is more reasonable than that of APWS. The analysis shows that APWS and APWS_RA schemes both have satisfying comprehensive performance, and APWS_RA is more effective.For the application of color reproduction for multispectral images, WF serial compression methods are designed through the proposed color perception error criterion. As a result, the multispectral image compression algorithms, which have advantages of low-complexity, good illuminant stability and supporting consistent color reproduction across devices, can be suggested by combining the excellent MSE(mean square error) based coding schemes. For example, in this thesis, WF_APWS_RA algorithm is designed through combining APWS_RA coding scheme in WF method. Experimental results indicate that under various illuminations, WF serial algorithms have obvious superiority on color accuracy.To be more effective in automatically removing the spatial-spectral redundancy, starting from the discussion of the spatial structural similarity of multispectral images, Optimal Leaders and Optimal Leaders-Color clustering algorithms are proposed, with spectral MSE and colorimetric error as their similarity measures respectively. Both of them are used to obtain multispectral images’ spatial sparse equivalence representation which can eliminate spatial redundancy adaptively. Based on spectral clustering, the universal composite multispectral image compression method named OLP-X is designed. Meanwhile, OLCP-X composite compression method is presented by applying colorimetric clustering. Furthermore, in combination of OLCP-X compression and WF coding idea, OLCPW-X is proposed to ensure both colorimetric accuracy and spectral reconstruction accuracy(X refers to traditional multispectral compression algorithms). In these three methods, spatial sparse equivalence representation coding and error compensation mechanism are employed, spectral dimensionality reduction and X compression are applied to further remove spectral and spatial redundancy for representative spectrums and predicted residual images respectively. The analysis illustrates that new algorithms supported by OLP-X and OLCP-X have better spectral reconstruction precisions than the original X schemes; OLCPW-X method obviously improves both rebuilt spectral and colorimetric precisions, and compared with OLCP-X supported algorithm, the algorithm supported by OLCPW-X gains better colorimetricprecision with comparative spectral accuracy.Finally, a compression algorithm for multispectral image, based on spectral dictionary learning and sparse representation, is proposed by exploiting data features of spectrum dimension. In the scheme, K-SVD algorithm is adopted for training a redundant dictionary from typical similar spectrum sets. The sparse representative coefficients of each spectrum vector are obtained by the dictionary, which can remove spectral redundancy of multispectral images. And then the equivalent nonzero sparse coefficients are quantified and stored. Experiments show that in comparison with classical principal component analysis, the new algorithm acquires better compression performance on spectrum dimension. Besides, the spectrum dictionary learning can also be combined with compressed sensing theory or spatial decorrelation technologies to further expand its application.