CN103177447B - A kind of hyperspectral image abnormal detection system based on FPGA - Google Patents

Abstract

The present invention relates to a kind of hyperspectral image abnormal detection system based on FPGA.The present invention includes data input module, covariance matrix solves module, generalized inverse solves module and result solves module, data input module is responsible for the equalization process of window matrix data, its output terminal connects the input end that covariance matrix solves module, the output terminal that covariance matrix solves module connects the input end that generalized inverse solves module, generalized inverse solves the generalized inverse that module is used for solving covariance matrix, the input end of its output terminal connection result computing module.The present invention can realize the high rate bioreactor of hyperspectral image data, and can upgrade to hardware as required, to realize more high performance requirements.

Description

A kind of hyperspectral image abnormal detection system based on FPGA

Technical field

The present invention relates to a kind of hyperspectral image abnormal detection system based on FPGA.

Background technology

Although high-spectrum remote sensing does well in wave band number and resolution, but its googol for requiring that ageing data processing causes very large obstacle according to amount and high data transfer rate, proposes many new efficient data processing methods process high spectrum image for solving this difficult problem scientific research personnel.

The fast development of high spectrum resolution remote sensing technique makes remote sensing image data magnanimity increase, the Data processing center station being located at earth can not process in time to the remote sensing image data of explosive increase, seriously hinders the effective of the valuable data of magnanimity and makes full use of.

And, owing to being subject to the restriction of transmission link bandwidth, high-spectrum remote sensing needs the lossy compression method of carrying out large ratio before being sent to ground, to such an extent as to ground is in the face of passing a large amount of detailed information losing image when image recovers back, causes serious image quality loss.Therefore EO-1 hyperion high-speed processing technology is the one necessity method addressed this problem, and especially on satellite, just can carry out real-time pre-service to remote sensing image data, increase substantially the utilization factor of data.Treated view data can carry out vast scale compression when reducing distortion as far as possible, alleviates the pressure that data transmission link bandwidth requirement is higher.

Along with the high speed development of EDA Technique, significantly improving of semiconductor technology makes FPGA scale increasing, and calculated performance is more and more stronger, and the Hyperspectral imagery processing algorithm that many traditional software methods realize is become hardware circuit realization of High Speed.

Summary of the invention

The object of the invention is counting yield in order to solve the PC when carrying out abnormality detection process to high spectrum image can not the problem of requirement of real time, proposes a kind of Hyperspectral imagery processing implement device based on FPGA.The high rate bioreactor of hyperspectral image data can be realized, and can upgrade to hardware as required, to realize more high performance requirements.

The technical scheme that technical solution problem of the present invention is taked:

It is a kind of that the hyperspectral image abnormal detection system based on FPGA comprises data input module, covariance matrix solves module, generalized inverse solves module and result solves module, data input module is responsible for the equalization process of window matrix data, its output terminal connects the input end that covariance matrix solves module, the output terminal that covariance matrix solves module connects the input end that generalized inverse solves module, generalized inverse solves the generalized inverse that module is used for solving covariance matrix, the input end of its output terminal connection result computing module.

Described data input module mainly comprises window matrix module, mean value computation module, matrix subtraction block and matrix transpose module.Window matrix module produces window matrix data, and the mean data that mean value computation module produces is input to matrix subtraction block, and matrix subtraction block data result is input to the transposition of matrix transpose module solution matrix.

Described covariance matrix solves module and mainly comprises matrix multiplication module and fixed point turns floating data module.Matrix multiplication result of calculation is input to fixed point and turns floating-point module and carry out data type conversion.

Described generalized inverse solves the nucleus module that module is whole system, mainly comprises generalized inverse and solves controller, Matrix Calculating order module, matrix multiplication module, generalized inverse computing module and data memory module composition.Data memory module is used for temporarily storing input matrix data, the matrix data that Matrix Calculating order module and matrix multiplication module read in data memory module carries out parallel computation, both results are input to generalized inverse computing module and carry out generalized inverse and solve, and generalized inverse solves the flowing of controller primary responsibility data dispatching and controls the action of other module.

Described result computing module mainly comprises mean vector computing module, and vector matrix dot product module and vector dot module, relate generally to vector dot computing.

Beneficial effect of the present invention:

(1) system have employed there is Parallel Computing Performance advantage FPGA as platform, and fully excavate the concurrency of computation process, finally be mapped as concrete hardware circuit to realize calculating process, improve system operations performance, compare traditional software method and substantially increase data processing speed.

(2) core calculations module have employed the accuracy that single precision data type fully ensure that result of calculation.

(3) FPGA has the dirigibility of height simultaneously, and FPGA has the ability of reshuffling only to be needed to revise configuration file and both can realize amendment to internal circuit and maintenance.System based on FPGA exploitation can be upgraded fast and maintain easily, and can save R&D costs and reduce R&D risk.

Accompanying drawing explanation

Fig. 1 RX algorithm flow chart;

Fig. 2 is system architecture schematic diagram of the present invention;

Fig. 3 is data input module design framework figure;

Fig. 4 is that covariance matrix solves modular design frame diagram;

Fig. 5 is that generalized inverse solves modular design frame diagram;

Fig. 6 is Matrix Calculating order modular design frame diagram;

Fig. 7 is result computing module design framework figure.

Embodiment

By reference to the accompanying drawings, the invention will be further described.

Be illustrated in figure 1 RX algorithm flow chart, can find out that this algorithm finds abnormal object by the RX operator value of data in the detection window of calculating local from single window RX algorithm principle.Mobility detect window carries out view picture high spectrum image detecting all abnormal objects that just can find in image.Therefore, we can think that whole algorithm carries out calculating in units of the window matrix that forms of the data in the detection window of local.

According to space size and the resolution of image, RX algorithm detection window is set to 11 × 11 pixel sizes, detection window can comprise 121 pixels, according to the three-dimensional matrice data in the known window of high-spectral data feature being 11 × 11 × Nband (Nband is selected wave band number).Three-dimensional matrice must be converted into two-dimensional matrix to calculate according to calculating needs, turn to by original matrix the two-dimensional data matrix that dimension is Nband × 121.Then, the Mean Matrix of calculation window matrix, deducts with data matrix the estimated value that matrix of consequence that Mean Matrix obtains removes to try to achieve background covariance matrix.Calculate the generalized inverse matrix of covariance matrix, then bring acquired results matrix into abnormality detection operator and calculate.The abnormality detection operator of trying to achieve and threshold value compare and judge whether measuring point to be checked is abnormal object, then carry out abnormality detection operator by moving window to view picture high spectrum image and calculate the abnormity point found out in image.

As shown in Figure 2, the present embodiment comprises data input module, covariance matrix solves module, generalized inverse solves module and result solves module, the output terminal of data input module connects the input end that covariance matrix solves module, the output terminal that covariance matrix solves module connects the input end that generalized inverse solves module, and generalized inverse solves the input end of the output terminal connection result computing module of module.The realization of modules is described below in conjunction with each figure:

(1) data input module design.

Fig. 3 gives the frame diagram of data input module, and window matrix generating module realizes carrying out Windowing process to original image data, i.e. generating window data matrix.Produce reading and writing data enable signal and read/write address signal by address generator, read the pel data in hyperspectral image data memory module, be stored in window matrix RAM.In window data matrix, each row represents a pixel, next by the mean vector of all pixel vectors included by Mean Matrix module calculation window, and produces corresponding Mean Matrix.Realized the local center process of data in window again by matrix subtraction block, produce new data matrix.Produce its transposed matrix by matrix transpose process, output to and calculate covariance matrix module.Control module then can realize micro-slip window and slide in whole image scene array, thus realizes carrying out abnormality detection process to all view data except border.

1-1 address generator module

The effect of address generator module is used to produce window matrix data, and view data is sequential storage in memory, and namely each pixel vector data address is continuous print.Because window matrix size is 11 × 11, so window matrix data is that (N is selected wave band number in N × 121, i.e. each pixel vector element number), address generator module produces reading and writing data enable and reading, writing address signal, is taken out and sends into data processing module by pixel vector data in window and carry out calculation process.The key that window matrix generates is to produce corresponding read-write enable signal and read/write address signal.Order takes out first element a11 of each pixel vector from image data memory, a12, a1n forms the first row of window matrix, in like manner take out second element a21 of each pixel vector, a22 ... a2n forms the second row of window matrix, so until the data temporary storage of taking out in all window matrixes is in window matrix window_ram.

Design level and vertical two counter x_cnt, y_cnt.X_cnt is that horizontal direction counting produces and makes the control signal of window level movement, y_cnt be that vertical direction counting produces the control signal making the movement of window vertical direction, and x_cnt and y_cnt comes together to provide the control signal of window movement.Window matrix is from left margin, first move horizontally calculating window matrix center pixel being carried out to one-time detection operator, then, when it moves to right boundary, turn back to left margin and move down a line and then repeat to move horizontally, until window slips over whole image-region.

A ROM memory module is customized with IP kernel, be initialized as a mould 120 counter rom_cnt, each pixel vector data that can be used in cycling among windows matrix, therefore the pixel sequential value in each window is that n_cnt=x_cnt+y_cnt+rom_cnt (determines the value of its x_cnt and y_cnt of window matrix that is determined, the change of rom_cnt can have access to the value of each pixel in window matrix), the first row element address such as reading window matrix data is N*n_cnt-N, N*n_cnt-(N-1), N*n_cnt-1 (now, n_cnt=1).In like manner can read all row of window matrix data and be stored in window matrix store Window_Ram, subsequent arithmetic module can read data and carry out further calculation process from storage.

1-2 mean value computation module

Mean value computation module is used for the estimated value of the average producing vector data in window matrix, and produces Mean Matrix further.For accelerating arithmetic speed, N number of parallel processing element (PE) is set and carries out the cumulative of each row element of window matrix respectively, be sent to parallel PE computing module carry out accumulation operations according to the window matrix data that sequentially reads of row, obtain each row element and value.Calculate acquired results, by data strobe switching sequence stored in result memory.Every a line can be obtained mean vector with value divided by window matrix size, then mean vector is carried out expand each elements extend of mean vector become and just constitute Mean Matrix data with a window matrix size row element of a size.

1-3 matrix subtraction block

Matrix subtraction block mainly realizes the subtraction of window matrix and Mean Matrix, and the matrix of differences of generation is used for calculating the value of covariance matrix.It is fairly simple that matrix subtraction block realizes, and designs a control module and is used for producing reading and writing enable signal and reading, writing address signal.When external signal inputs enable matrix subtraction block, control module starts and produces to be read enable signal accordingly and reads address signal, reads data feeding computing module and carry out computing from window matrix and Mean Matrix memory module.Arrange a counter in the control module carry out counting operation to data calculating process thus produce computing end signal, as the enable signal of next module.

(2) covariance matrix module

Due to when carrying out Generalized Inverse Matrix and solving, its data are all that decimal and house calculation module contain the operation of a large amount of multiply accumulating, and fixed-point data is owing to being subject to the restriction of its precision and dynamic range not competent.Floating data can provide higher resolution in larger dynamic range, and the system that can ensure has higher precision.Fig. 4 gives the frame diagram that covariance matrix solves module, therefore covariance matrix realizes by a module separately because calculating is larger, major calculations is matrix multiplication operation, call multiply-accumulator concurrent operation realization matrix multiplication, add before output certain point number according to type turn single-precision floating-point data type block realize data conversion to provide subsequent module to carry out computing.

(3) generalized inverse solves module

Fig. 5 gives the frame diagram that Generalized Inverse Matrix solves module.The compositions such as controller, Matrix Calculating order module, matrix multiplication module, generalized inverse computing module and data memory module are solved primarily of generalized inverse.Generalized inverse solves the operation that controller is responsible for control data flowing and is coordinated each sub-function module, and wherein matrix multiplication calculating and rank of matrix calculate parallel carrying out, and effectively can improve computing velocity.Startup generalized inverse computing module after two modules have all calculated, this module mainly completes based on the generalized inverse interative computation process of mark method solution matrix until obtain Generalized Inverse Matrix.

3-1 solves generalized inverse mark method

Matrix Calculating generalized inverse module is the most complicated part of whole algoritic module, if known matrix A _{m × n}order be r, the generalized inverse algorithm of Matrix Calculating is as follows:

Step (1) calculates B=A ^ta;

Step (2) makes C ₁=I, initialization Matrix C _i;

Step (3) calculates $C_{i+1}=\frac{1}{i}tr\left(C_{i}B\right)I-C_i{A}^T;$

Step (4) calculates $A^{+}=\frac{r}{tr\left(C_{i}B\right)}{C}_i{A}^T;$ Note, C _i+1b=O, tr (C _ib) ≠ 0.

Ask the mark method of Generalized Inverse Matrix to need the order of known matrix, the Applying Elementary Row Operations of matrix can determine rank of matrix.The Applying Elementary Row Operations of matrix has three kinds:

A () exchanges two row;

B () takes advantage of all elements of certain a line with several k (k ≠ 0);

C () gets on the element that the k of certain a line all elements is doubly added to another row corresponding.

Three kinds of Applying Elementary Row Operations are combined, any non-zero matrix can be turned to Hermite standard form, wherein (1) and (2) plant conversion be easy to realize, but (3) kind ratio of transformation is more loaded down with trivial details, and object is that all the other elements of certain element column in matrix are turned to 0 entirely.Plant conversion for (3), have employed a kind of rectangle diagonal line computing method.By analyzing, can find that this algorithm has very strong concurrency, very applicable FPGA realizes.

3-2 Generalized Inverse Matrix solves modular design

3-2-1 Matrix Calculating order modular design

Shown in Fig. 6, its top-level module mainly comprises Matrix Calculating order controller, data input, data storage, interface module, pivoting module, pivot turns to 1 (row times multiplication), row exchanges and data export 8 sub-function module.Matrix Calculating order module is the nucleus module during generalized inverse solves, and the order r calculated is the important parameter that generalized inverse solves module, controls the iterative computation number of times of generalized inverse calculating sub module.Specific implementation is as follows:

The data of matrix to be solved are inputted pivoting submodule according to row order by step (1), then by first row, choose principal element (in a certain row first nonzero element), if principal element when line number equals the row of cycle calculations number of times, does not need to exchange, all elements of being expert at by principal element exchanges to the row (now calculation times counter remains unchanged) that line number is calculation times;

Step (2) computation process starts, first principal element place row data are passed through the value of divider computing divided by principal element successively, principal element becomes 1, other element is updated to the multiple of principal element, be equivalent to doubly take advantage of operation to matrix one row element, each element is multiplied by the value of coefficient 1/ principal element.

Step (3) makes other element of principal element column be directly 0, principal element the transformed value of other outer row element of being expert at then saved by 3-1 described in rectangle diagonal line computing method upgrade, carry out calculation times counter when renewal completes and add 1;

Step (4) repeats (1), (2) and (3) calculating process until proceed to last row, last computation process does not need by (1), (2) and (3) step, can by directly judging to obtain a result.Whether detailed process is last element of first last row of judgment matrix is zero, if zero rank of matrix equals calculation times, if not zero moment rank of matrix equals calculation times add one, then export the result after calculating settling signal and module arithmetic and further calculate to input subsequent module.

The operation of whole module is carried out in order under state machine controls, whole calculating process is decomposed into several subprocess carried out in certain sequence to carry out computing realization, and subprocess just can represent by corresponding states, subprocess carry out order calculate process just corresponding states conversion process.Therefore whole computing engineering can be realized by the thought of state machine.

First under reset signal rst_n controls, it is IDLE state that state machine carries out homing action init state, then under input control signal function, state machine starts the S1 that gets the hang of, now correspond to pivoting module and carry out work, select the principal element (principal element is not equal to 0) in certain column data of input.When pivoting module completes pivoting operation, state machine enters the operation (doubly taking advantage of operation in Applying Elementary Row Operations) that NextState S2 carries out principal element to become 1, the inverse of principal element is doubly multiplied by all elements that pivot is expert at.Complete after doubly taking advantage of operation, then carry out changing judgement, judge whether to need to carry out row swap operation, when needs swapping rows, the state machine S3 that gets the hang of carries out row swap operation: when not needing row to exchange, state machine directly get the hang of S4 to except principal element be expert at except other element carry out renewal rewards theory (data processing).State S3 is capable, and the S4 that gets the hang of equally when having exchanged carries out data processing operation.

After last round of Data Update terminates, calculate round counter and add one, after S4 state completes, judge that whether calculation times is equal with required calculating total degree, when calculation times is less than total degree state machine from S4 state proceed to S1 state restart next round Data Update operation, to the last required matrix is turned to Hermite standard form matrix, then the state machine IDLE that gets the hang of waits for that next matrix computations starts.

3-2-2 cycle calculations modular design

Start working obtaining rank of matrix r Posterior circle computing module, this module comprises Matrix Calculating mark, matrix subtraction, matrix multiplication operation.Ask trace of a matrix can realize by carrying out diagonal entry adding up, matrix multiplication can call front matrix multiplication module used, and matrix subtraction realizes fairly simple, only needs the reading matrix data be stored in RAM to input a subtracter and calculates in order.Rank of matrix is that a key parameter is just embodied in this module as previously mentioned, upgrades Matrix C _iloop calculation in rank of matrix r directly control the number of times of cycle calculations.

First under reset signal rst_n controls, it is IDLE state that state machine carries out homing action init state, and then under input control signal function, state machine starts the S1 that gets the hang of, and now corresponds to and asks matrix multiplication module to carry out work, ask input matrix C _iwith the product matrix C of B _ib.When matrix multiplication module completes multiply operation, state machine enters NextState S2 to carry out the computing asking trace of a matrix (then carrying out accumulation operations with being taken out by matrix diagonals line element), and just the inverse of cycle calculations number of times is doubly multiplied by C _ithe all elements of B matrix.After completing and asking trace of a matrix to operate, then enter the state machine S3 that gets the hang of and carry out the operation of matrix subtraction.

When one takes turns after cycle calculations terminates, calculate round counter and add 1, after S3 state of operation completes, judge that whether calculation times is equal with required calculating total degree (rank of matrix r), when calculation times is less than total degree, state machine proceeds to S1 state from S3 state and restarts new round loop computing function, to the last tries to achieve Matrix C _r.Again by Matrix C _rsend into result computing module to carry out next step and calculate, then the state machine IDLE that gets the hang of waits for that next matrix computations starts.

(4) result computing module design.

Fig. 7 gives the design framework figure of result computing module, and result computing module controller is responsible for the generalized inverse matrix of covariance matrix to be input to computing module, and reads the data vector of window matrix center pixel.Computation of mean values vector submodule calculates the average of the every a line of window matrix to produce mean vector.The difference of calculation window matrix center pel data vector and mean vector, then carry out point multiplication operation with the generalized inverse matrix of covariance matrix, result vector carries out vector dot computing with window matrix center pel data vector with the difference value vector of mean vector again can obtain net result.Single Precision Floating Point Multiplier Based is called in vector dot computing and subtracter composition floating multiplication summing elements realizes.

Claims (1)

1. the hyperspectral image abnormal detection system based on FPGA, comprise data input module, covariance matrix solves module, generalized inverse solves module and result solves module, it is characterized in that: data input module is responsible for the equalization process of window matrix data, its output terminal connects the input end that covariance matrix solves module, the output terminal that covariance matrix solves module connects the input end that generalized inverse solves module, generalized inverse solves the generalized inverse that module is used for solving covariance matrix, the input end of its output terminal connection result computing module;

Described data input module comprises window matrix module, mean value computation module, matrix subtraction block and matrix transpose module; Window matrix module produces window matrix data, and the mean data that mean value computation module produces is input to matrix subtraction block, and matrix subtraction block data result is input to the transposition of matrix transpose module solution matrix;

Described covariance matrix solves module and comprises matrix multiplication module and fixed point turns floating data module; Matrix multiplication result of calculation is input to fixed point and turns floating-point module and carry out data type conversion;

Described generalized inverse solves the nucleus module that module is whole system, comprises generalized inverse and solves controller, Matrix Calculating order module, matrix multiplication module, generalized inverse computing module and data memory module composition; Data memory module is used for temporarily storing input matrix data, the matrix data that Matrix Calculating order module and matrix multiplication module read in data memory module carries out parallel computation, both results are input to generalized inverse computing module and carry out generalized inverse and solve, and generalized inverse solves the operation that controller is responsible for control data flowing and is coordinated each sub-function module;

Described result computing module comprises mean vector computing module, and vector matrix dot product module and vector dot module, relate generally to vector dot computing.