Hyperspectral band selection and extraction method and system based on FPGA (Field Programmable Gate Array)

Abstract

The invention discloses a hyperspectral band selection and extraction system and method based on an FPGA (Field Programmable Gate Array), and mainly solves the problems of high complexity, poor real-time performance, low compressed sensing band selection efficiency and more occupied resources of the existing band selection algorithm. According to the implementation scheme, the method comprises the following steps of: generating data storage and an address by using a three-layer nested counter, generating a storage address of a hyperspectral image, reading an external original hyperspectral image according to the address, and storing the external original hyperspectral image into an RAM (Random Access Memory) for caching; setting an n-order linear feedback shift register, feeding back a polynomial and a seed value, generating a sparse random binary matrix, and linearly outputting a 0-1 random number to a subsequent operation module; and carrying out operation on the input original hyperspectral data and the sparse random binary by adopting a judgment accumulation mode to obtain selected hyperspectral data. According to the method, the number of wavebands needing to select dimension reduction can be flexibly adjusted, internal resource consumption is reasonably reduced, the timeliness of waveband selection in a satellite-borne scene is improved, and the method can be used for efficient spectral information processing of satellite-borne, airborne and other resource-limited platforms.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and further relates to a hyperspectral band selection and extraction method and system, which can be used for efficient spectral information processing on resource-constrained platforms such as spaceborne and airborne platforms. Background Art

[0002] Hyperspectral images have the characteristic of "combining spectrum and image", and their spectral resolution can reach the nanometer level. Therefore, the rich spatial and spectral data with a large amount of information obtained by hyperspectral remote sensing has a wide range of applications and prospects in many fields such as geological exploration, agricultural production, food safety, environmental detection, military applications, and medicine. A typical hyperspectral image exists in the form of a three-dimensional data cube, including the spatial dimension and the spectral dimension. However, with the continuous development of hyperspectral imager technology, the hyperspectral image has been significantly improved in both spatial resolution and spectral resolution. However, this often means a heavy burden on computing and storage. Especially on actual airborne and spaceborne platforms equipped with hyperspectral imaging and data processing equipment, it is difficult to quickly process such a huge amount of data due to limited resources. Secondly, due to the continuity of spectral bands in hyperspectral data, the different spectral bands are highly correlated and the band information redundancy is serious. In fact, it is not necessary to use and store all band information during processing. Finally, during the process of obtaining hyperspectral data, affected by various factors such as atmospheric scattering, a large amount of noise will appear in some bands, resulting in almost unusable data for these bands. In the subsequent image processing and analysis process, it is necessary to perform necessary elimination processing on these bands, otherwise it will have an adverse impact on the image processing results. Therefore, improving the real-time related technology of hyperspectral data is the core breakthrough point of hyperspectral remote sensing technology.

[0003] Most traditional signal compressions use the Nyquist sampling method, which requires the sampling frequency to be maintained at twice the highest frequency. For high-frequency signals or signals with a wide bandwidth, this will result in an extremely high sampling frequency, thus generating a large amount of data. For example, in hyperspectral imaging, each pixel point contains data of multiple bands, making the overall data volume very large, posing a huge challenge to storage and transmission.

[0004] The patent document with the application number CN202310416474.3 discloses "A Hyperspectral Band Selection Method Based on Spatial-Spectral Dual Attention Branches". Its implementation scheme is as follows: (1) Preprocess the hyperspectral image dataset to construct an unlabeled hyperspectral image training sample subset; (2) Construct a spatial-spectral dual attention branch network to fuse feature maps to obtain a spatial-spectral attention feature map; (3) Construct a reconstruction network model to reconstruct the hyperspectral image; (4) Optimize the training of the spatial-spectral dual attention branch network by reducing the error between the reconstructed hyperspectral image and the original hyperspectral image; (5) Construct a composite waveband subset selection index and select the top-ranked wavebands as the waveband subset; (6) Test the performance of the selected waveband subset. Due to its high complexity and the fact that model training and network optimization during the operation process consume excessive computing resources and computing time, this method cannot meet the requirements for hyperspectral image band selection in real-time scenarios. Summary of the Invention

[0005] The object of the present invention is to propose a low-complexity hyperspectral band selection method and system based on FPGA in view of the deficiencies in the above situation, so as to greatly reduce computing resources and computing time and achieve real-time selection of hyperspectral image bands.

[0006] Based on a new sampling method for sparse signals - compressive sensing CS proposed by Terence Tao, David Donoho and others, the present invention utilizes the sparsity characteristics of signals to achieve efficient undersampling and reconstruction through non-adaptive linear projection. Its implementation scheme includes:

[0007] 1. A hyperspectral image band selection method based on FPGA, characterized by comprising the following steps:

[0008] (1) Use a three-layer nested counter to generate data storage and addresses, and generate the storage address of the hyperspectral image. Read the external original hyperspectral image according to this address and store it in the RAM for caching.

[0009] (2) Set an n-order linear feedback shift register, feedback polynomial and seed value to generate a sparse random binary matrix, and linearly output 0-1 random numbers to the subsequent operation module.

[0010] (3) Perform operations on the input original hyperspectral data and the sparse random binary by means of judgment and accumulation to obtain the downsampled hyperspectral data, that is, the selected hyperspectral data.

[0011] (4) Store the selected hyperspectral data in the random access memory RAM inside the FPGA.

[0012] 2. A hyperspectral image band selection system based on FPGA, characterized by comprising

[0013] The perception matrix dynamic generation module: used to generate a sparse binary matrix and output an element sequence to control the operation logic;

[0014] The compression operation module is used to multiply the input original hyperspectral data by a sparse random binary matrix to obtain the dimension-reduced hyperspectral data;

[0015] The control module is used to coordinate the data flow and calculation timing through three sub-modules to control the states of the above-mentioned modules;

[0016] The storage management module: used to generate a three-layer nested counter for data storage and addresses, generate the storage addresses of hyperspectral images, read the external original hyperspectral images according to these addresses and store them in the RAM for caching, and store the dimension-reduced hyperspectral data.

[0017] 3. An electronic device, characterized in that it includes: a processor and a memory, the memory stores original hyperspectral data and hyperspectral data with band selection completed; the processor is used to run the program instructions to execute the steps of the above-mentioned hyperspectral image band selection method.

[0018] 4. A processor-readable storage medium, characterized in that the read storage medium stores FPGA instructions, and the FPGA instructions are used to cause the FPGA to execute the steps of the above-mentioned FPGA-based hyperspectral compressive sensing band selection method.

[0019] Compared with the prior art, the present invention has the following advantages:

[0020] First, aiming at the problems of time occupied by complex algorithms and a large amount of computing resources encountered by traditional hyperspectral image band selection algorithms, the present invention utilizes the sparsity of the hyperspectral data set and adopts a universal model based on compressive sensing for band selection, simplifies the operation process, improves the operation speed and greatly reduces the resource consumption of the algorithm.

[0021] Second, aiming at the problems of excessive additional storage and computing resource consumption during the selection and generation of sparse matrices in compressive sensing, the present invention adopts a linear feedback shift register method to generate a sparse random binary matrix, reducing the pressure of storing the sparse random binary matrix.

[0022] Third, aiming at the problem of low processing efficiency and poor stability caused by data translation during hyperspectral endmember extraction in existing implementation platforms, the present invention designs an FPGA hardware system for this band selection algorithm. With the acceleration effect of the hardware platform, the time consumption is greatly reduced; the band selection system can meet the real-time requirements in the spaceborne scenario.

[0023] The simulation results show that for hyperspectral images with various spatial sizes and spectral dimensions in different scenarios, the system designed by the present invention can flexibly adjust the initial parameters, significantly increase the algorithm rate on the premise of ensuring that the selected hyperspectral data still has relevant properties, and highly control the consumption of hardware resources. Description of the Drawings

[0024] Figure 1 Flowchart of the hyperspectral image band selection method provided in Embodiment 1 of the present invention;

[0025] Figure 2 Block diagram of the hyperspectral image band selection system provided in Embodiment 2 of the present invention;

[0026] Figure 3 Block diagram of the electronic equipment structure provided by the present invention;

[0027] Figure 4 is the true value map of the hyperspectral image;

[0028] Figure 5 is the simulation result of target detection using the present invention and the prior art;

[0029] Figure 6 is the simulation result of endmember extraction using the present invention and the prior art.

[0030] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0031] It should be noted that the step numbers in the specification and claims of the present invention are only for clear description of the implementation schemes of the present invention for easy understanding, and their sequence numbers are not limited.

[0032] Embodiment 1, Hyperspectral Image Band Selection Method Based on FPGA

[0033] Referring to Figure 1 , the implementation steps of this example are as follows:

[0034] Step 1, read in the hyperspectral image.

[0035] In this embodiment, the internal storage address of the hyperspectral image in the FPGA is generated by using a three-layer nested counter, and the externally input hyperspectral image is read into the block random access memory (BRAM) in the FPGA for caching in the address order.

[0036] The three-layer nested counters are a band counter, a column counter, and a row counter respectively. When the band counter is full once, the column counter increments by one. And so on, when the column counter is full once, the row counter increments by one. Through the nested counting method of the three counters, the BRAM storage address within the image slice is obtained.

[0037] The BRAM is a dedicated random storage block resource integrated within the FPGA. By using this storage block resource, immediate reading of the image can be achieved.

[0038] In an embodiment of the present invention, a hyperspectral image of TI_TE with a size of 200*200*189 is read in.

[0039] Step 2, generate a sparse random binary matrix.

[0040] 2.1) Set an n-order linear feedback shift register, a feedback polynomial, and a seed value to generate a sparse random binary matrix:

[0041] 2.1.1) Design the characteristic polynomial and the number of registers of the linear feedback shift register according to the data volume of the hyperspectral image, and connect these registers in series and preset their initial states;

[0042] 2.1.2) According to the selected seed number, characteristic polynomial, and number of registers, perform exclusive OR on the bits specified by the characteristic polynomial in this series of connected registers and input the result of the exclusive OR into the first bit of this series of registers, then shift the remaining data in the registers one bit to the right, and output a random number through the shift of the last register;

[0043] 2.1.3) Repeat step 2.1.2), continuously generate 0-1 random number elements of the sparse random binary matrix, and generate 2 N -1 random numbers according to the preset N registers, and form a sparse random binary matrix with random numbers that meet the number of elements required for the hyperspectral data.

[0044] In this example, the size of the TI_TE hyperspectral dataset is 200*200*189. According to this hyperspectral data size, a 20-order linear feedback shift register is used, the preset seed number is 8H in hexadecimal, and the characteristic polynomial is as follows.

[0045] y = x 20 +x 17 +1

[0046] According to the selected characteristic polynomial above, the registers involved in the exclusive OR operation in this example are the 20th and the 17th. After the exclusive OR operation, the exclusive OR result will be output to the first register in series, and then the data retained in the remaining registers will be shifted one bit to the right, and a 0-1 random number will be output from the 20th register. According to the number of selected registers, the random number period that can be generated in this example is 1,048,575, which meets the requirement of 800,000 random numbers for the TI_TE hyperspectral dataset with the number of downsampling bands up to 20.

[0047] 2.2) Linearly output the 0-1 random numbers generated by the linear feedback shift register to the subsequent compressive sensing band selection operation.

[0048] Step 3, perform the compressive sensing band selection operation.

[0049] 3.1) Calculate the hyperspectral signal r after band selection according to the sensing matrix Φ composed of the original hyperspectral signal r and the 0-1 random numbers generated by the feedback shift register Φ :

[0050] r Φ = Φr + n

[0051] where n represents noise, which can be ignored in actual operations.

[0052] Although the calculation process of this hyperspectral signal utilizes the sparse characteristics of the hyperspectral dataset, it can avoid the complex reconstruction process of compressive sensing. Compared with the traditional band selection algorithm, it greatly reduces the overall computational complexity of the algorithm. However, since it does not utilize the computational characteristics of the hardware platform, the matrix multiplication operation can be further optimized by using the method of judgment and accumulation to further improve the computational speed of this algorithm after hardware deployment.

[0053] 3.2) Optimize the algorithm of the hardware platform for the matrix multiplication operation of the input original hyperspectral data and the sparse random binary matrix by using the method of judgment and accumulation:

[0054] 3.2.1) Use a selector to judge whether the elements of the sensing matrix are 0 or 1, and execute different calculation processes respectively:

[0055] When the element of the sensing matrix is 0, the hyperspectral element at this time is discarded;

[0056] When the element of the sensing matrix is 1, the hyperspectral element is retained and an accumulation operation is performed using an adder. After judging the elements of the number of original hyperspectral bands, the multiplication operation of one pixel and one row of the sensing matrix is completed;

[0057] 3.2.2) Repeat step 3.2.1) until all pixels complete the matrix multiplication operation.

[0058] When performing operations in this way of judgment and accumulation, since only one dual-channel data selector and one accumulator are required, it avoids using a multiplier that is slower and consumes more resources. It can improve the overall speed of the algorithm while fully utilizing the data characteristics of the sensing matrix and reduce resource consumption.

[0059] Step 4, perform data storage

[0060] Store the hyperspectral data after compressive sensing, that is, the hyperspectral data with band selection completed, in the random access memory (RAM) within the FPGA to complete the calculation of hyperspectral image band selection.

[0061] Embodiment 2, Hyperspectral Image Band Selection System Based on FPGA

[0062] Refer to Figure 2 , this example includes: a sensing matrix dynamic generation module 1, a compression operation module 2, a control module 3, and a module 4 storage management module, where: the sensing matrix dynamic generation module 1 includes a preset sub-module 11, an exclusive OR operation sub-module 12, and an iterative sub-module 13; the compression operation module 2 includes a judgment sub-module 21, an accumulation sub-module 22, and a counting sub-module 23; the control module 3 includes an idle sub-module 31, a compressive sensing call sub-module 32, and a termination sub-module 33;

[0063] The working principle of the entire system is as follows

[0064] The sensing matrix dynamic generation module 1 is used to generate a sparse binary matrix. Through the preset sub-module 11, it designs the characteristic polynomial and the number of registers of the linear feedback shift register, and connects these registers in series and presets the initial state. The specific parameters it designs are based on the specific size of the hyperspectral image. The designed characteristic polynomial determines the number and positions of the registers for the exclusive OR operation in the exclusive OR operation sub-module 12; the exclusive OR operation sub-module 12 is used to generate a one-bit 0-1 random number, that is, by performing an exclusive OR on the number and positions of the registers determined by the preset sub-module 11 and inputting the result of the exclusive OR into the first bit of this string of registers, and then shifting the data retained in the remaining registers one bit to the right, and outputting a random number through the last register shifted, to complete the output of a 0-1 random number; the iterative sub-module 13 is used to repeatedly call and run the exclusive OR operation sub-module 12 to continuously generate the 0-1 random number elements of the sparse random binary matrix, and generate 2 N -1 random numbers according to the N registers set by the preset module 11, and form a sparse random binary matrix with random numbers that meet the number of elements of the sparse random binary matrix required for hyperspectral data, and linearly transmit it to the compression operation module 2 in the generation order.

[0065] The compression operation module 2 is used to perform compressive sensing operations on hyperspectral data using a hyperspectral universal model, that is, multiplying the generated sensing matrix by the original hyperspectral data. Through the judgment sub-module 21, it uses a selector to judge whether the elements of the sensing matrix linearly output by the sensing matrix dynamic generation module 1 are 0 or 1, and respectively executes different calculation processes:

[0066] When the element of the sensing matrix is 0, the hyperspectral element at this time is discarded;

[0067] When the element of the sensing matrix is 1, the data is sent to the accumulation sub-module 22. When the sensing matrix is 1, this hyperspectral element is accumulated by an adder in the accumulation sub-module 22, and the counting module 23 is used to count and judge the number of accumulated elements according to the number of original hyperspectral bands: if the corresponding counting number is reached, the multiplication operation of one pixel and one row of the sensing matrix is completed; otherwise, the element accumulation continues; during this process, taking advantage of the parallel processing of the hardware platform FPGA, multiple parallel processing units PE are called to perform the above judgment and accumulation work to replace the single-line processing process of software. While making full use of the characteristics and resources of the FPGA itself, the operation speed is greatly improved.

[0068] The control module 3 is used to mobilize the data streams of the sensing matrix dynamic generation module 1 and the compression operation module 2 when selecting hyperspectral bands. That is, when the start signal is pulled high by the idle sub-module 31, all registers in the system are initialized and configured. After synchronously completing the reset and loading of operation parameters, it automatically jumps to the compressive sensing call sub-module 32, which reads hyperspectral data from the memory band by band and sends it to the compression operation module 2, and outputs the operation result data to the RAM, and then clears the accumulator. After all pixels are completed with compression conversion, it switches to the termination sub-module 33. The termination sub-module 33 resets the control signals of all operation units, sends an operation completion flag signal, and then automatically returns to the idle sub-module to prepare for the next round of calculation tasks.

[0069] The storage management module 4 is used to generate a three-layer nested counter for data storage and addresses, generate the storage address of the hyperspectral image, read the external original hyperspectral image according to this address and store it in the RAM for caching, and generate an address according to the flag signal of the control module 3 to store the hyperspectral data after the compression operation module 2 completes the compressive sensing band selection.

[0070] It should be noted that the counting sub-module 23 can design a counter circuit inside the FPGA to be used to judge and call the accumulation sub-module 22 according to the number of bands of the original hyperspectral data set to complete the compression operation. It can also be used to judge between different pixels when storing pixel-by-pixel hyperspectral data.

[0071] Refer toFigure 3 , the present invention also provides an electronic device, which includes a control bus, a processor, and a memory. All program instructions and cached hyperspectral image data can be stored in the memory, and the processor implements the method steps in the previous embodiments by running the program instructions in the memory.

[0072] The processor can be implemented in the form of a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present specification.

[0073] The memory can be implemented in the form of a read-only memory (ROM), a random-access memory (RAM), a static storage device, a dynamic storage device, etc. The memory can store program instructions and other application programs. When implementing the technical solutions provided in the embodiments of the present specification through software or firmware, the relevant program codes are stored in the memory and are called and executed by the processor.

[0074] The bus includes a data transmission path for transmitting information between various components of the device, such as the memory and the processor.

[0075] The processor-readable storage medium provided by the present invention stores FPGA instructions, and the FPGA instructions are used to cause the FPGA to execute all steps of the hyperspectral compressive sensing band selection method based on the FPGA in this embodiment.

[0076] The processor-readable media used in this embodiment include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be processor-readable instructions, data structures, program modules, or other data. The carriers of the processor storage medium in this example include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage, or other magnetic storage devices, or any other non-transmission media, and these storage carriers are used to store all information and data that can be processed.

[0077] It should be noted that although the above device only shows the processor, the memory, and the data bus, in the specific implementation process, the device also includes other components that can operate normally.

[0078] The effects of the present invention can be further illustrated by the following simulation experiments:

[0079] 1. Simulation conditions

[0080] The simulation and deployment are carried out on the hardware platform Vivado, and the model of the FPGA development board used is Virtex-7 XC7VX690T ffg1157-2. The preprocessing of hyperspectral data will be processed on the computer platform.

[0081] The data used in the simulation are TI and TE simulation data. The spatial resolution of the data is 200*200, and there are 189 bands in total. It contains five-row and five-column targets. The targets are composed of the true spectral signals of five substances, namely Alunite, Buddingtonite, Calcite, Kaolinite, and Muscovite, extracted from cuprite data, and the background is composed of the true spectrum of a certain background pixel extracted from cuprite data. The targets in the same row are of roughly the same substance but different in size, and the targets in different rows are of different substances; the targets in the first column are pure spectral targets with a size of 4*4 pixels, the second column are pure spectral targets with a size of 2*2 pixels, the third column are mixed spectral targets with 50% of each of the two substances and a size of 2*2 pixels, the fourth column are mixed spectral targets with 50% of the 1*1 target and 50% of the background spectrum, and the fifth column are mixed spectral targets with 25% of the 1*1 target and 75% of the background, as Figure 4 shown

[0082] II. Simulation content

[0083] Simulation 1: Under the above conditions, a simulation of target detection is carried out using the dataset after band selection of the present invention and the original hyperspectral dataset with the Automatic Target Generation Program (ATGP) algorithm. The results are as Figure 5 , where Figure 5 a is the detection result and running time using the dataset after band selection of the present invention, Figure 5 b is the detection result and running time using the original hyperspectral dataset.

[0084] From Figure 5 it can be seen that the detection results obtained by using the ATGP algorithm on the original hyperspectral image of the dataset after compressive sensing band selection of the present invention are the same, but the present invention has a great advantage in the detection operation time. That is, the original data has 189 bands, and the number of bands after band selection of the present invention is only 20, which greatly reduces the operation time and improves the overall detection effect.

[0085] Simulation 2: Under the above conditions, a simulation of endmember extraction is carried out using the dataset after band selection of the present invention and the original hyperspectral dataset with the Simplex Growing Algorithm (SGA). The results are as Figure 6 , where Figure 6 a is the extraction result and running time using the dataset after band selection of the present invention, Figure 6 b is the extraction result and running time using the original hyperspectral dataset.

[0086] From Figure 6 It can be seen that the extraction results obtained by using the SGA algorithm on the original hyperspectral image of the dataset after compressive sensing band selection in the present invention are the same, but the present invention has great advantages in the extraction operation time. That is, the original data has 189 bands, and the number of bands after band selection by the present invention is only 20, which greatly reduces the operation time and improves the overall effect of endmember extraction.

Claims

1. A method for hyperspectral image band selection based on FPGA, characterized in that, Including: 1) Use a three - layer nested counter to generate data storage and addresses, and generate the storage address of the hyperspectral image. According to this address, read the external original hyperspectral image and store it in the RAM for caching; 2) Set an n - order linear feedback shift register, feedback polynomial, and seed value to generate a sparse random binary matrix, and linearly output 0 - 1 random numbers to the subsequent operation module; 3) Perform an operation on the input original hyperspectral data and the sparse random binary by means of judgment and accumulation to obtain the dimension - reduced hyperspectral data, that is, the selected hyperspectral data; 4) Store the selected hyperspectral data in the random access memory (RAM) within the FPGA.

2. The method according to claim 1, wherein In step 2), by setting the n - order linear feedback shift register, feedback polynomial, and seed value, a sparse random binary matrix is generated, and its implementation includes the following: 2a) Design the characteristic polynomial and the number of registers of the linear feedback shift register, connect these registers in series, and preset the initial state; 2b) Perform exclusive - OR operations on some bits in this series of connected registers, input the result of the exclusive - OR into the first bit of this series of registers, then shift the remaining data in the registers one bit to the right, and output a random number through the last register; 2c) Repeat step 2b), continuously generate 0-1 random number elements of the sparse random binary matrix, and generate 2 N -1 random numbers according to the preset N registers, and form a sparse random binary matrix with random numbers that meet the number of elements of the sparse random binary matrix required for hyperspectral data.

3. The method according to claim 1, characterized in that In step 3), perform matrix multiplication operations on the input original hyperspectral data and the sparse random binary by means of judgment and accumulation, and its implementation includes: 3a) Use a selector to determine whether the element of the sensing matrix is 0 or 1, and perform different calculation processes respectively: When the element of the sensing matrix is 0, discard the hyperspectral element at this time; When the element of the sensing matrix is 1, retain this hyperspectral element and perform an accumulation operation using an adder. After judging all the elements of the original hyperspectral bands, complete the multiplication operation of one pixel and one row of the sensing matrix; 3b) Repeat step 3a) until all pixels complete the matrix multiplication operation.

4. A high - spectral image band selection system based on FPGA, characterized in that, Including The sensing matrix dynamic generation module: used to generate a sparse binary matrix and output an element sequence to control the operation logic; The compression operation module, used to perform matrix multiplication on the input original hyperspectral data and the sparse random binary to obtain the dimension - reduced hyperspectral data; The control module, used to coordinate the data flow and calculation timing through three sub - modules to control the states of the above - mentioned modules; The storage management module: used to generate a three - layer nested counter for data storage and addresses, generate the storage address of the hyperspectral image, read the external original hyperspectral image according to this address and store it in the RAM for caching, and store the dimension - reduced hyperspectral data.

5. The system according to claim 4, wherein The sensing matrix dynamic generation module includes: A linear feedback shift register, used to generate a sparse binary matrix, and its implementation process is as follows: A preset sub - module, used to design the characteristic polynomial and the number of registers of the linear feedback shift register, connect these registers in series, and preset the initial state; An exclusive - OR operation sub - module, used to perform exclusive - OR operations on some bits in this series of connected registers, input the result of the exclusive - OR into the first bit of this series of registers, then shift the remaining data in the registers one bit to the right, and output a random number through the last register; An iterator sub-module, which is used to repeatedly run the exclusive OR module to continuously generate 0-1 random number elements of the sparse random binary matrix, and generate 2 N -1 random numbers, and form a sparse random binary matrix with random numbers that meet the number of elements of the sparse random binary matrix required for hyperspectral data.

6. The system according to claim 4, wherein The compression operation module includes: A judgment sub-module, which is used to judge whether the elements of the sensing matrix are 0 or 1 by using a selector, so as to execute different calculation processes respectively: When the element of the sensing matrix is 0, the hyperspectral element at this time is discarded; When the element of the sensing matrix is 1, the data is sent to the accumulation sub-module; The accumulation sub-module is used to accumulate this hyperspectral element by using an adder when the sensing matrix is 1. After judging the elements of the original hyperspectral band number, the multiplication operation of one pixel and one row of the sensing matrix is completed; The counting sub-module is used to call the accumulation sub-module according to the band number of the original hyperspectral data set to complete the compression operation.

7. The system according to claim 4, wherein The control module includes: An idle sub-module, which is used to initialize the register configuration when the received start signal is pulled high. After synchronously completing the reset and loading of the operation parameters, it automatically jumps to the compressive sensing sub-module; The compressive sensing call sub-module is used to read the hyperspectral data from the memory band by band and send it to the operation module, output the operation result data to the RAM, and clear the accumulator. After all pixels complete the compression conversion, the state machine is switched to the termination sub-module. The termination sub-module. It is used to perform two key operations: first, reset the control signals of all operation units and send an operation completion flag signal; second, automatically return to the idle sub-module to prepare for the next round of calculation tasks.

8. The system according to claim 4, wherein The compression operation module multiplies the input original hyperspectral data with the sparse random binary matrix. By taking advantage of the parallel processing of the FPGA, multiple parallel processing units (PEs) complete the matrix multiplication operation by means of judgment and accumulation.

9. An electronic device, characterized in that, It includes: A processor and a memory, The memory stores the original hyperspectral data and the hyperspectral data with band selection completed; The processor is used to run the program instructions to perform the sensing matrix generation or compressive sensing operation as described in claims 1 to 3.

10. A processor-readable storage medium, characterized in that, The read storage medium stores FPGA instructions, and the FPGA instructions are used to make the FPGA execute the FPGA-based hyperspectral compressive sensing band selection method described in claims 1 to 3.

Images