Research On Hyperspectral Image Compression And Implement On DSP

A hyper-spectral imaging sensor can get image data with very narrow bandwidth over the same spatial area, by analysis the captured images; we can get one-dimensional spectral information of detected target. Trough that, we will easy to identify the target. The hyper-spectral image has great application value both in use by public and military. Because it has several or hundreds of spectral bands, the hyper-spectral image has enormous data quantity. Data compression can significantly reduce hyper-spectral data volumes to more manageable size for storage, communication and process.At first, a combined method based on improved PCA and integer wavelet transform is proposed for hyperspectral image compression in this paper. Principal component analysis (PCA) can effectively reduce the spectral correlation of hyperspectral image and integer wavelet transform by using lift scheme is widely used for spatial decorrelation. According to the code speed have dramatically decrease when the spatial size become large. The hyperspectral images are partitioned into several blocks with same size and each block is encoded by PCA and integer wavelet transform independently. A non-linear model is setup to estimate the optimal retained numbers of PCs at any compression ratio. When the optimized compression methods are using on the hyperspectral images of the AVIRIS instrument and our developing hyperspectral imager, the compression effects is competitive and it runs fast comparing with common PCA followed by integer wavelet transform.Through analysis and comparison of current various algorithms, a mixed compression algorithm based on prediction and integer wavelet transform is proposed to realize in hardware. First, an optimal inter-band predictor is designed for de-correlating the spectral redundancy of hyperspectral image. Next, its spatial redundancy is removed with an efficient integer wavelet transform. Then, an improved embedded zero-tree wavelet (EZW) encoder is employed for encoding the obtained transformed coefficients. For the necessity to compress the image data cube in real time from the hyperspectral imager on space and/or air platform, this paper adopts a high-powered Digital Signal Processor (DSP) of TMS320DM642 to realize the proposed mixed algorithm. Since the data volume of spectral image is so huge that it is impossible to be processed in the internal memory of DM642 one-off, it is partitioned into several blocks of same size and then each of them is encoded by the internal CPU. The effective means to solve the problem is to use DMA and CACHE for high speed transportation and read of the image data in the memory. Through modifying the mixed algorithm and optimizing its algorithmic language, the processing efficiency of the program was significantly improved, compared the non-optimized one. The experiment show that the mixed algorithm based on DSP runs much faster than the algorithm on personal computer. The work of this article has founded a real-time hyperspectral image compression platform.