CN1290056C - Ultra-spectrum image real-time compression system based on noise masking - Google Patents

Abstract

A hyperspectral image real-time compression system based on a noise masking algorithm, comprising a host, a high-speed image slave processor, and an image acquisition device, the image acquisition device is connected to the slave processor through a CAB bus interface and connected to the host through a PCI bus; the The main frame is connected bidirectionally with the slave processor through the PCI bus, and the host is respectively connected to the image display and the image data compression result memory, and is characterized in that: a. The image noise masking compression module and the host interface of the dual bond are provided on the slave processor , so as to accept the command data from the host computer and transmit the image real-time compression result data to the host computer; b. The host computer control template is provided on the host computer, from the processor interface, display module and storage module, the control module accepts the original image data and the user command to form a system data source, and send commands and receive image real-time compression result data from the slave processor to the slave processor; the host control module sends the original image data to the display module and the image real-time compression result to the storage module to form a compressed file.

Description

Real-time Compression System of Hyperspectral Image Based on Noise Masking

technical field

The present invention relates to a hyperspectral image real-time compression system, in particular to an image compression system based on noise analysis, which minimizes the influence of noise introduced by compression algorithms on image information by estimating the signal-to-noise ratio of real-time hyperspectral images , realizes the high-fidelity compression of the hyperspectral image, and also guarantees the image compression rate of the image compression system.

Background technique

There have been many reports on the research results of hyperspectral image compression, but most of them stay in the stage of theoretical research and software processing, and there are not many real-time hardware applications. The reasons are as follows: firstly, the data volume of hyperspectral images is huge, which requires high processing power of hardware, and it was difficult to complete it in real time with hardware circuits in the past; It is closely combined with the development of imaging spectrometers and is rarely used in practice.

With the development of large-scale integrated circuit technology, many integrated circuit chips with high speed, high reliability and low power consumption have been developed, and real-time processing of digital images by hardware has become possible. The real-time compression system composed of multiple high-performance processing chips combined with the hyperspectral image compression algorithm has reached the practical stage. The miniaturization requirements of imaging spectrometers and the pressure of storage and transmission of a large number of hyperspectral image data also make the research and development departments of imaging spectrometers begin to pay attention to research work in this area. For example, the Offshore Imager (COIS) carried by the NEMO naval map observation satellite designed by the US Naval Research Laboratory includes two imagers, including 60 and 150 spectral bands with a width of 10nm. The data rate of these two imagers reaches 145Mb/s. In order to achieve a data compression rate of 10 times (lossy) and reduce the burden of hyperspectral images stored on the star and transmitted to the ground, the U.S. Naval Research Laboratory has developed a real-time processing system for feature extraction and data compression on the star, called Optical real-time adaptive spectral discrimination system (ORASIS). In the experiment conducted by unmanned aircraft, the actual compression effect achieved by ORASIS is: when the compression ratio = 12, the peak signal-to-noise ratio = 48.0db; when the compression ratio = 20, the peak signal-to-noise ratio = 44.2db. In China, Beijing University of Aeronautics and Astronautics, National University of Defense Technology, Tsinghua University, Harbin Institute of Technology and other units have also done related research. The imaging spectrometer carried by the "Aerospace Tsinghua-1" microsatellite jointly developed by China's Tsinghua University, China Aerospace Machinery and Electronics Corporation and the University of Surrey consists of green (0.52-0.6μm), red (0.63-0.69μm), near-infrared ( 0.76-0.94μm) composed of three bands, using the a-MPBTC (Adaptive Moment Preserving Block Truncation Coding) compression method to achieve a 4-fold compression rate.

However, the actual remote sensing images inevitably have the interference of random noise. The shortcoming of analyzing these imaging spectral image compression systems is that they do not consider the noise characteristics of the image. When the image is compressed by a certain multiple, the image edge The original noise of the transition zone is amplified by the system, and spreads and pollutes the high-bit data of the image, resulting in the loss of image information. In short, when these compression systems are used to compress images, the image information cannot be reasonably preserved.

Contents of the invention

As mentioned above, the technical problem to be solved by the present invention is how to increase the compression factor of the image without increasing the loss of image information according to the noise pollution of the hyperspectral image. Therefore, the object of the present invention is to provide a kind of hyperspectral image compression system, through the noise analysis to each band data of original image, design effective image compression method, carry out targeted compression to each band data of hyperspectral image, establish based on Noise-masked hyperspectral image real-time compression system.

Technical solution of the present invention is as follows:

According to a kind of hyperspectral image real-time compression system based on noise masking algorithm of the present invention, comprise host computer, high-speed image slave processor, and image acquisition device, described image acquisition device is connected with high-speed image slave processor through CAB bus interface and via The PCI bus is connected to the host; the host is bidirectionally connected to the high-speed image slave processor through the PCI bus, and is also connected to the image display and the image data compression result memory respectively. The characteristics are:

a. The high-speed image slave processor is provided with an image noise masking compression module connected with a bidirectional data flow and a host interface, and the image noise masking compression module includes an original image input unit, an image signal-to-noise ratio estimation unit, and an image signal-to-noise ratio estimation unit connected in sequence. A sub-block segmentation unit, an image sub-block two-dimensional quantized cosine transform unit, an image signal-to-noise ratio reconstruction unit, and an image data coding compression unit, the image signal-to-noise ratio reconstruction unit has a reconstructed image with a signal-to-noise ratio that is higher than that of the original image The signal-to-noise ratio is 10db larger, so as to accept the command data from the host and transmit the image real-time compression result data to the host;

b. The host is equipped with a host control module, a slave processor interface, a display module, and a storage module. The control module accepts the original image data sent by the image acquisition device and the commands sent by the host to form a data source of the system, and passes through Compress the result data when sending commands from the processor interface to the high-speed image slave processor and accept the image sent by compression; the host control module sends the original image data to the display module and sends the image real-time compressed result data to the storage module to form a compressed file and store it in the image Data compression result memory.

Further, the host is a portable PC;

Described high-speed image is made of multiple DSP chip image processing boards from processor;

The image sub-block dividing unit divides the original image into 4×4 or 5×5 or 8×8 sub-blocks.

The hyperspectral image compression system based on the noise masking algorithm of the present invention is characterized in that: for the designed hyperspectral image compression algorithm, it can be implemented with a hardware system without modification or as little modification as possible, saving the work required for re-evaluating the algorithm quantity.

The system of the present invention also has strong flexibility, and the system can easily improve the algorithm according to the actual application requirements without re-developing and designing the entire hardware system, and has scalability and portability, and the hardware design is convenient for future expansion of the system Or apply to a new platform.

Because the present invention uses the compression method based on the noise masking effect, the signal-to-noise ratio of the reconstructed image is controlled to be greater than the original image signal-to-noise ratio by 10db, so that the noise introduced by the compression algorithm is covered by the original image noise, thereby ensuring Improved the image compression factor.

In a nutshell, the present invention makes full use of the noise characteristics of hyperspectral images, minimizes the impact of noise introduced by compression algorithms on image information, realizes high-fidelity compression of hyperspectral images, and ensures image compression rate , compared with the traditional hyperspectral image compression system, it has obvious technical progress and substantial significance.

Description of drawings

FIG. 1 is a schematic diagram of the hardware architecture of the present invention.

Fig. 2 is a schematic diagram of the software architecture of the present invention.

Fig. 3 is a schematic diagram of the time-sharing parallel processing method of multiple DSPs on the image processing board of the present invention.

Fig. 4 is a schematic diagram of the workflow of the single-chip DSP two-stage double-buffer pipeline of the present invention.

Fig. 5-1 is a schematic structural diagram of the image compression module in the noise-based masking algorithm of the present invention.

Fig. 5-2 is a flowchart of the compression process of the noise masking compression method of the present invention.

Fig. 6 is a schematic diagram of the experimental principle of the noise masking effect of the present invention.

Fig. 7 is a graph showing the experimental results of the noise masking effect of the present invention.

Fig. 8-1 and Fig. 8-2 are graphs of application examples of the compression method based on noise masking of the present invention, respectively.

Detailed ways

A better embodiment of the present invention is provided below according to Fig. 1～Fig. 8-2, and further provides the technical detail of the present invention in conjunction with the description of this embodiment, so that can better understand structural feature and function of the present invention Effect.

Referring to FIG. 1, the hardware architecture of the present invention includes a system control host 10, a high-speed image slave processor 20, an image acquisition board 22, a CAB bus interface 21, and the like. Wherein: the system control host 10 adopts a portable PC, and the high-speed image slave processor 20 adopts a multi-chip TMS320C6201-based image processing board Python/C6 processing board. Its task of described host 10 is to control the whole system, utilize its friendly interface to receive user commands, start/stop the work of the slave processor, that is, the high-speed image slave processor 20, change the working program of the DSP on the slave processing board, etc., and collect The relevant image data and image compression information are displayed in real time, and the compressed data is stored. In addition, the software of the host computer 10 is also responsible for decompressing and playing back the compressed results during post-processing after the real-time work is completed. The task of the Python/C6 processing board is to accept the image data, let the DSP process it according to the program selected by the host, and output the compressed result. The original spectral image data collected by the image acquisition board 22 is divided into two transmissions, one is directly input into Python/C6 through the CAB bus interface 21 for processing; one is transmitted to the host computer 10 through the computer PCI bus 30 for real-time compression and is compressed by the image display 12 displays for easy monitoring. The data compression result of Python/C6 is transmitted to the host computer 10 through the PCI bus 30, and is stored or further processed by the memory 11.

Referring to Fig. 2, the software architecture of the present invention is divided into two parts: a host computer 10 platform (Host) and a processor 20 platform (Slave). The main modules on the host computer 10 include a main control module 101 , a slave processor interface module 102 , an image display module 103 and a data storage module 104 . Its main function is to provide a good human-machine interface, accept user commands, control the operation of the slave processor machine 20, receive the compression result of the microprocessor board, and realize the real-time display of the original image and the storage function of the compression result by using technologies such as multithreading . The main modules of the software of the slave processor 20 include an image compression algorithm module 201 and a host interface module 202 . Its main function is to use a variety of code optimization methods on the DSP platform, and use TI's high-efficiency code compiler to give full play to the hardware resource performance of the DSP, and complete the real-time data compression work according to the user's command and transmit it to the host. The entire software starts with raw images and user commands as data sources and ends with compressed result files.

As shown in Figure 3, it is a schematic diagram of the multi-chip DSP time-sharing parallel processing method on the Python/C6 image processing board of the present invention. Each piece of DSP in the parallel processing mode independently processes one channel of data to complete a complete compression algorithm task. When all channels of data are processed, all modules are output at the same time. The advantage of this structure is that the task allocation is simple, the functions and algorithms of each DSP are the same, and it is easy to realize. In the system of the present invention, in combination with the characteristics of multi-DSP image processing boards, the parallel processing method of time-sharing pipeline is adopted, that is, the input hyperspectral image data is divided into a certain size according to the size of the buffer memory configured by each DSP. The frames are piped to each DSP according to time sequence, and the processing results of each DSP are also combined according to time sequence and transmitted to the host computer 10 through the PCI bus.

The DSP selected in this embodiment is the TMS320C6201 chip of TI Company, and the CPU can access the IDRAM twice in one clock cycle, which is much more efficient than the external memory that needs to expand the memory interface and external bus through the EMIF. In order to give full play to the performance of DSP, it is necessary to use IDRAM as a buffer for data exchange. However, for image compression that needs to process a large amount of data, the capacity of IDRAM is too small to meet the needs of fast data input and output. To this end, a two-level cache structure is adopted: off-chip memory SDRAM (16MB) is used as a frame buffer, and IDRAM (64KB) is used as an image sub-block cache. The original image data is first input and temporarily stored in the frame buffer, and then sent to the on-chip sub-block cache in units of sub-blocks for processing by the CPU. The DMA controller of C6201 has 4 independent programming channels, allowing DMA transmission of 4 different contents. For memory-intensive image compression systems, I/O throughput is critical to the system's processing power and efficiency. Using the four DMAs of C6201 on Python/C6, two-level I/O buffering can be realized, that is, the exchange of image data from the bus to the relatively slow SDRAM memory of C6201, and the exchange of SDRAM memory to the on-chip high-speed data RAM of C6201. In this way, the data compression process can be carried out in the on-chip high-speed data RAM, and the throughput of the data is greatly improved.

As shown in Figure 4, it is a schematic diagram of the workflow of a single-chip DSP two-stage double-buffer pipeline. Use DMA2 and DMA3 to realize the operation of the first-level input device to the input frame buffer BufFrameIn and the output frame buffer BufFrameOut to the output device respectively, and use DMA0 and DMA1 to realize the operation of the second-level input frame buffer BufFrameIn to the on-chip sub-block buffer BufBlockIn and on-chip sub-block respectively Buffer BufBlockOut to output framebuffer BufFrameOut operations. These two stages both utilize the ping-pong cache, and use the pipeline parallel processing method, so that the input and output and data compression processing are performed simultaneously. The dotted line and the solid line in the figure are two mutually exclusive operations, which cannot be performed at the same time, but the operations represented by the same line can be performed at the same time. Operations indicated by dotted lines and solid lines are executed alternately to form a pipeline. The specific pipeline parallel processing process is as follows (assuming that the initialization filling of the two-stage pipeline has been completed):

1) Start DMA2 and collect the next frame of image data from the input device to BufFrameIn1;

2) Start DMA3, output the last frame image compression result from BufFrameOut1 to PCI bus;

3) start DMA0, and the next image sub-block in this frame image data BufFrameIn2 is sent to BufBlockIn2;

4) Start DMA1, and transfer the compression result of the last image sub-block from BufBlockOut2 to BufFrameOut2;

5) Compress and process the image sub-block in BufBlockIn1, and put the result into BufBlockOut1;

6) Wait until the transfer of DMA0 and DMA1 ends, repeat steps 3, 4, and 5 and alternately use the two on-chip sub-block caches until all the frames are compressed

7) Wait until the transfer of DMA2 and DMA3 ends, repeat steps 1, 2, 3, 4, 5, and 6 and alternately use the extended frame buffer to start compressing subsequent frames.

Such a processing process makes the data transmission of the 4 DMA channels and the compression work of the DSP core completely parallel, and ensures that the main calculation tasks are carried out in the on-chip cache, fully exerting the performance of the system hardware and improving the data throughput.

As shown in Figure 5-1 and Figure 5-2, the hyperspectral image compression method in this embodiment makes full use of the noise characteristics of the image. Before compressing the image, the signal-to-noise ratio of the image must first be estimated. The basic idea of this system for image signal-to-noise ratio estimation is: because it is difficult to select a uniform area of a certain size, the image is divided into small areas one by one, and these small areas can basically be considered uniform; The LSD (Local Standard Deviation, local standard deviation) is used as the local noise size, and the LSD of the interval with the largest total number is selected as the average noise value of the entire image. The specific operation steps are as follows:

Divide the image into 4×4, or 5×5,…, or 8×8 small blocks. For each image sub-block, the LM (Local Mean) of the signal is obtained by the following formula:

$\mathrm{<span\; class="notranslate">\; LM</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

Here, Si is the gray value of the i-th pixel in the image sub-block; N is the total number of all pixels in the image sub-block. LSD is obtained by the following formula:

$\mathrm{<span\; class="notranslate">\; LSD</span>}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; [</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; LM</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ]</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}$

For uniform image sub-blocks, LSD is small, while for non-uniform image sub-blocks, such as sub-blocks containing image edges or texture features, LSD is large. Calculate the LM of the entire image (denoted as LMo), the LSD of all image sub-blocks, and find the largest and smallest LSD of all image sub-blocks.

Between the minimum and maximum LSD, several intervals of equal intervals are established. Arrange the LSDs of all sub-blocks into corresponding intervals in sequence according to the value. The number of LSDs in each interval is counted, and the average value of the LSDs in the interval with the largest count value is the noise of the entire image, which is denoted as LSDo.

The signal-to-noise ratio (SNR) of the entire image can be obtained from the following formula:

$\mathrm{<span\; class="notranslate">\; SNR</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 20</span>}\mathrm{<span\; class="notranslate">\; log</span>}\frac{{\mathrm{<span\; class="notranslate">\; LM</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}}{{\mathrm{<span\; class="notranslate">\; LSD</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}}$

That is, its basic steps are: step 51, the original image is input to the original image input unit 2011; step 52, the image signal-to-noise ratio estimation unit 2012 estimates the signal-to-noise ratio of the input image; step 53, the image sub-block segmentation unit 2013 divides the image Be divided into 8×8 image sub-blocks one by one; Step 54, image sub-block two-dimensional quantized cosine transform unit 2014 carries out two-dimensional discrete cosine transform (2DCT) to each image sub-block; Step 55, image signal-to-noise ratio reconstruction unit 2015 according to The SNR estimation result quantifies the transformed data, so that the SNR of the reconstructed image is 10db higher than the SNR of the original image; step 56, the image data encoding and compression unit 2016 performs Huffman encoding on the quantized data to realize image data compression.

Refer to FIG. 6 for a schematic diagram of the experimental principle of the noise masking effect. Use an ideal noise-free image as the original input signal, and add zero-mean Gaussian white noise with different variances to it as the noisy input signal; use the JPEG compression algorithm based on DCT to perform CR times on the input noisy signal Compression; decompress the compressed data, and get the reconstructed image. Refer to FIG. 7 for a graph showing the experimental results of the noise masking effect. In the figure, the abscissa is σ _c ² /σ _i ² , which is the ratio of the quantization distortion variance to the input noise variance, and the ordinate is σ ² /σ _i ² , which is the ratio of the total noise variance of the reconstructed image to the input noise variance, as shown in As shown, the three curves respectively represent the relationship curves when the input signal-to-noise ratio is 50db, 35db, and 20db. It can be seen from the figure that when σ _c ² /σ _i ² <-10db, the curves under the three different input noises start to tend to 0 at the same time, that is, σ _i ² ≈σ ² (the total noise variance of the reconstructed image is approximately equal to the input noise variance), This shows that the quality of the reconstructed image depends almost on the magnitude of the input noise and has nothing to do with the compression noise (quantization noise). Therefore, the selection result of the best compression noise is also obtained: as long as the compression noise (quantization noise) is controlled to be 10db lower than the input noise, that is, σ _c ² /σ _i ² <-10db, the reconstruction image can be reconstructed without increasing distortion In the case of increased data compression.

The compression method based on noise masking of the present invention is realized on the basis of σ _c ² /σ _i ² <-10db. Give an example of OMIS (Modular Imaging Spectrometer Practical) hyperspectral image compression. For the noisy OMIS hyperspectral image, the signal-to-noise ratio (SNR) of the original image is:

${\mathrm{<span\; class="notranslate">\; SNR</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; lg</span>}\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

The reconstructed image signal-to-noise ratio SNRo is:

${\mathrm{<span\; class="notranslate">\; SNR</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; lg</span>}\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

According to the experimental results of noise masking effect, let

${{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\frac{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; 10</span>}}$

Substitute into SNRo, get

${\mathrm{<span\; class="notranslate">\; SNR</span>}}_{\mathrm{<span\; class="notranslate">\; o</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; lg</span>}\frac{{\mathrm{<span\; class="notranslate">\; 10</span>}\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="10"\; label="lg"\; filenames="C20041002569100141.png,C20041002569100142.png"\; state="\{\{state\}\}">lg</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; lg</span>}\frac{{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}{{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; SNR</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}$

Therefore, for the lossy compression of noisy hyperspectral images, as long as the signal-to-noise ratio (SNRo) of the reconstructed image is controlled to be greater than the SNR of the original image (SNRi) by more than 10db, the compression factor of the image can be increased without increasing the distortion of the reconstructed image. Based on this idea, different degrees of lossy compression are performed on the data of each band of the OMIS image, and the size of SNR is controlled by changing the quantization distortion of the compression algorithm, so that each band SNR _o =10+SNR _i . Figure 8-1 and Figure 8-2 show the application example curves of the compression method based on noise masking. Among them, Figure 8-1 is the signal-to-noise ratio curve of the original image of each band of OMIS and the reconstructed image; Figure 8-2 is the curve of image compression ratio of each band of OMIS. It can be seen from the figure that the image data of any band has been compressed to the greatest extent according to its noise pollution, and the reconstructed image has basically no loss of image information because its SNRo is always 10db larger than the SNRi of the original image. Got great fidelity.

Claims (4)

1, a kind of hyperspectral image real-time compression system based on noise masking algorithm, comprise main frame (10), high-speed image from processing machine (20), and image acquisition device (22), this image acquisition device (22) is through CAB bus interface (21) be connected with from processor (20) and be connected with main frame (10) through PCI bus (30); (10) connect image display (12) and image data compression result memory (11) respectively, it is characterized in that: A. be provided with image noise masking compression module (201) and host interface (202) that become two-way connection on processor (20), this image noise masking compression module (201) comprises the original image input unit (2011) that connects in sequence ), an image signal-to-noise ratio estimation unit (2012), an image sub-block segmentation unit (2013), an image sub-block two-dimensional quantized cosine transform unit (2014), an image signal-to-noise ratio reconstruction unit (2015), and an image data coding compression unit ( 2016), the reconstructed image signal-to-noise ratio of the image signal-to-noise ratio reconstruction unit (2015) is 10db larger than the original image signal-to-noise ratio, so as to accept the command data from the host computer (10) and transmit the image real-time compression result data to host(10); B. be provided with host control template (101) on host computer (10), from processor interface (102) and display module (103) and storage module (104), this control module (101) accepts original image data and user order And form system data source, and send command and accept image real-time compression result data from processor interface (102) to slave processor (20); Host control module (101) sends original image data to display module (103) and will The image real-time compression result is sent to the storage module (104) to form a compressed file. 2. The hyperspectral image real-time compression system based on noise masking algorithm according to claim 1, characterized in that: said host (10) is a portable PC. 3. The hyperspectral image real-time compression system based on noise masking algorithm according to claim 1, characterized in that: said slave processor (20) is composed of multiple DSP chip image processing boards. 4. The hyperspectral image real-time compression system based on noise masking algorithm according to claim 1, characterized in that: said image sub-block segmentation unit divides the original image into 4×4 or 5×5 or 8×8 sub-blocks piece.

Images