CN106556831A - GRADIENT PROJECTION METHODS based on FPGA - Google Patents

Abstract

The invention discloses a kind of GRADIENT PROJECTION METHODS based on FPGA, mainly solution prior art execution cycle is long, and resource utilization is low, the problem of time delay length.Implementation step is：1. the output data of radar imagery is solved using GRADIENT PROJECTION METHODS；2. in Vidado High Level Synthesis HLS softwares to the solution procedure using GRADIENT PROJECTION METHODS to carrying out many suboptimization；3. the GRADIENT PROJECTION METHODS after optimization is generated into the IP core with radar imagery function by Method at Register Transfer Level RTL；4. the IP core of generation is called in FPGA, according to radar sequential in the input input measured data of IP core, the output data of radar imagery is obtained in the output end of IP core, is realized the radar imagery in FPGA.The present invention can shorten the construction cycle by original algorithm rapid deployment on FPGA, reduce time delay, improve resource utilization, can be used for forward sight microwave imaging.

Description

GRADIENT PROJECTION METHODS based on FPGA

Technical field

The invention belongs to technical field of image processing, more particularly to a kind of GRADIENT PROJECTION METHODS, can be used for forward sight microwave into Picture.

Background technology

Gradient projection GPSR algorithms are the cores of forward sight microwave imaging, based on the sparse signal that compressed sensing CS is theoretical Reconstruct is most time-consuming committed step in gradient project algorithms, be all matrix the step of realize sparse reconstruct with vector, vector with Vector, the multiplication between matrix and matrix calculate operation, when the data of matrix-vector are excessive, bring larger pressure to storage Power, it is huge that the amount of calculation calculating time also becomes.It is longer to inevitably result in operation time when being operated on CPU, is unfavorable for compiling Journey, designer wish to possess a kind of low delay, and the processor of high operation efficiency is realizing the algorithm.

Apply within 2015 good just in thesis《The design of forword-looking imaging semi-matter simulating system and realization based on GPU》In relate to And how gradient projection is realized using GPU, though this method realizes concurrent operation, execution cycle is shortened, but still up to not To the requirement of low delay, and GPU work box body products are big not readily portable so that the application in missile-borne is extremely inconvenient, and GPU Power dissipation ratio it is larger.

The content of the invention

Present invention aims to the deficiency of above-mentioned prior art, proposes a kind of gradient projection side based on FPGA Method, to reduce time delay, reduces power consumption, it is easy to using carrying.

The technical scheme is that what is be achieved in that：

One. know-why

FPGA as programmable application specific processor, from structure for be highly-parallel, possess lower work(than CPU Consumption and more powerful computing capability.FPGA trades space for time in algorithm design, is exactly by increasing hardware resource and hard Part design cost, the lifting implementation for exchanging processing speed and performance for are hardware-accelerated.

FPGA hardware design is transferred to C language from the other transmitting stage of RTL register stages, using C language by high-level comprehensive Direct creation of knowledge property right core is closed, then algorithm is completed to combine by calling the IP core.It is high-level using Vivado Comprehensive HLS, writes the subsystem of the IP core of Xilinx matches SEL, so that it may simpler using C language or C++ high-level languages The clean parallel performance for quickly having given play to FPGA.Present invention virtex7 Series FPGAs realize the forward sight of gradient project algorithms into The method of picture, is optimized operation to which, is realized with completing the high efficiency to gradient projection.

2nd, technical scheme

According to above-mentioned principle, the present invention to realize that step includes as follows：

(1) obtain the echo data s of the target of actual measurement^uWith pattern data F_θ, wherein θ ∈ (1,10000), u ∈ (1, 196)；

(2) calculate echo data s^u=F_θσ^uIn scene scatters factor sigma to be restored^u：

(2a) the first weighted factor β ∈ (0,1) of search and the second weighted factor are setSelection is initially set out Point z₀, and counter is set to 0；

(2b) calculate initial step length：a₀∈[α_min,α_max],

Wherein α_minRepresent and limit a₀The minimum of a value of scope, α_maxRepresent and limit a₀The maximum of scope, and 0<α_min<α_max, g_k It is the gradient direction of projection,The transposition of T representing matrixs；

(2c) under conditions of the first weighted factor β and the second weighted factor μ is met, find the kth that following formula is set up Secondary iteration step length a_k：

WhereinRepresent First-order Gradient value, ()₊Expression takes the positive portion of (), z_kIt is feasible during expression kth time iteration Point；

(2d) calculate new feasible pointz_k+1Represent feasible point during+1 iteration of kth；

(2e) convergence detection is carried out to feasible point, judges z_k+1Value whether meet following end condition：WhereinIt is normal number, tolP is a minimum value parameter, if meeting the termination Condition, then export z_k+1Value, as scene scatters factor sigma to be restored^u；If be unsatisfactory for, k=k+1 is made, return to step (2b)；

(3) step (2) is optimized in the Vivado HLS softwares that FPGA is carried, the field to be restored after being optimized Scape scattering coefficient

(4), after the completion of optimizing, Method at Register Transfer Level RTL is converted into imaging function in Vivado HLS softwares IP core, calls the IP core in FPGA, to optimization after scene scatters coefficient to be restoredAgain arranged Arrange and pass through vector superposed, finally give the two dimensional image of scene to be restored, complete the gradient projection based on FPGA.

The present invention has advantages below compared with prior art：

1) present invention is optimized to the GRADIENT PROJECTION METHODS in Vivado HLS softwares, reduces time delay, here mistake In journey, developer reduces the workload of developer without the need for grasping the knowledge with regard to register transfer rank RTL.

2) present invention with Vivado HLS softwares with derived from the form of IP-Catalog with gradient projection forward sight into As the universal IP cores of function, virtex7 Series FPGAs are not only applicable to, other serial FPGA are applied also for, Carrying easy to use.

3) due to being deployed to original algorithm on FPGA rapidly using Vivado HLS softwares, thus the present invention has The workload for reducing hardware programming of effect, improve projection realizes efficiency, reduces power consumption, shortens the construction cycle.

Description of the drawings

Fig. 1 be the present invention realize general flow chart.

Fig. 2 is to optimize the data flow dataflow pattern diagram adopted by GRADIENT PROJECTION METHODS in the present invention.

Fig. 3 is the simulation result figure that forword-looking imaging is realized with the present invention.

Specific embodiment

With reference to Fig. 1, the present invention's realizes that step is as follows：

Step 1. solves scene scatters coefficient to be restored according to the echo data and pattern data of the target of actual measurement.

(1a) set up the Mathematical Modeling of gradient projection：

If degree of rarefication is N for the length of primary signal x of K, calculation matrix y is obtained by projection matrix A：

Y=Ax<1>

Using Lagrange multiplier by formula<1>It is expressed as：

Due to l₁Norm non-differentiability, according to gradient projection need primary signal x is resolved into into positive number u and negative v two Point, i.e. v_i=(- x_i)₊, u_i=(x_i)₊, i=1 ...., N；

By formula<2>It is converted intoAnd z >=0<3>

Wherein, the transposition of T representing matrixs,

λ represents regularization parameter, b=A^TY,

By formula<3>Obtain the First-order Gradient value of object function F (z)：

(1b) calculate initial step length a₀：

Wherein g_kThe gradient direction of expression projection, 0<α_min<α_max, α_minRepresent and limit a₀The minimum of a value of scope, α_maxRepresent limit Determine a₀The maximum of scope；

(1c) the first weighted factor β ∈ (0,1) of search and the second weighted factor are setSelection is initially set out Point z₀, counter is set to 0；

(1d), in (1c) is met under conditions of the first weighted factor β and the second weighted factor μ, find kth time iteration step Long a_kMake formula<6>Set up：

Obtain new feasible pointFirst-order Gradient value during kth time iteration is represented, z_kRepresent feasible point during kth time iteration, z_k+1Represent feasible point during+1 iteration of kth；

(1e) convergence detection is carried out to feasible point, judges z_k+1Value whether meet formula<7>End condition,

If meeting end condition, z is exported_k+1Value, k values are updated if being unsatisfactory for, k=k+1 is made, are returned to (1b) continue executing with, whereinIt is normal number, tolP is a minimum value parameter；

(1f) according to (1a) to (1e), the echo data of the target to surveying carries out pulse compression, obtains scene to be restored Scattering coefficient σ^u：

I.e.

Wherein t_mThe slow time is represented,Represent the scattering coefficient of p-th unit, and u ∈ (1,196), P ∈ (1,10000), θ_p And β_pAzimuth and the angle of pitch of relative radar, formula are represented respectively<9>InThe echo-signal measured value of certain range cell is represented, Corresponding to formula<1>In y values, F is the pattern function with orientation time change, corresponding to formula<1>A in formula, σ^uIt is to treat extensive The scattering coefficient of multiple scene, corresponding to the z in (1e)_k+1；

To formula<9>Sparse constraint condition is added, scene scatters factor sigma to be restored is obtained^u

By formula<10>It is converted into solution minimum l₁The optimization problem of norm, i.e., thrown using gradient when signal reconstruction is carried out Image method is to formula<10>Solved, solved scene scatters factor sigma to be restored^uValue.

Step 2. is optimized to step 1, the scene scatters coefficient to be restored after being optimizedThis step it is concrete Realization is carried out by the Vivado HLS softwares that FPGA is carried, and its process is as follows：

(2a) select to launch instruction set_directive_unroll in the Vivado HLS softwares that FPGA is carried, will follow Ring sentence launches, and realization completes circulate operation within the shorter clock cycle；

(2b) array is processed in Vivado HLS softwares, selects array split order array_partition, Complete is selected in type, 1 is selected in dimension, realization is loaded into all data within a clock cycle, shortens Clock cycle；

(2c) operation of data flow dataflow is carried out in Vivado HLS softwares, select data flow operations instruction set_ Directive_dataflow, is this advantage of parallel device using FPGA, realizes executed in parallel, its operation principle such as Fig. 2 institutes Show, wherein Fig. 2 (a) is represented and do not done the execution sequence before optimizing, and Fig. 2 (b) represents holding after optimizing using data flow dataflow Row order；

If not using data flow dataflow pattern, gradient projection program presses four function fun_1 in Fig. 2 (a), Fun_2, fun_3, fun_4 are that order is performed, i.e., fun_1 functions start to perform fun_2 functions after the completion of performing, and have performed Perform fun_3 functions after fun_2 functions again, performed fun_3 functions and performed fun_4 functions again；

Only need to etc. after previous function performs a clock cycle in data flow dataflow pattern, next function is just Can start to perform, gradient projection program presses four function fun_1 in Fig. 2 (b), fun_2 executed in parallel, as long as fun_1 functions Start to perform fun_3 functions by when having output, while fun_2 functions begin to perform fun_4 functions when having output；

Scene scatters coefficient to be restored after being optimized after the completion of execution

Step 3. generates IP core.

After the completion of optimization, by input/output bus in Vivado HLS softwares, IPCatalog is selected, after optimizing Containing scene scatters coefficient to be restoredGRADIENT PROJECTION METHODS, generated with imaging function by Method at Register Transfer Level RTL IP core.

Step 4. calls IP core in FPGA.

(4a) in the clock signal sys_clk rising edge of IP core, when commencing signal ap_start is high level, It is input into the echo data targ_data and pattern data dire_data of the target of actual measurement to IP core simultaneously；

(4b), after a clock cycle, when idle signal ap_idle is low level, show IP core just in work Make；As idle signal ap_idle and when completing signal ap_done and being high level, show that IP core has completed work, Now output data ap_return of IP core is final scene scatters coefficient to be restoredTreat restoration scenario Scattering coefficientCarry out vector superposed, finally give the two dimensional image of scene to be restored.

The effect of the present invention can be further embodied by emulation experiment.

1. experiment condition：

If there is a point target in scene, radar works for forward sight, does not consider the impact of range migration；

2. experiment content：

Radar foresight imaging is carried out to a point target with the method for the present invention, as a result such as Fig. 3, wherein Fig. 3 (a) is represented and bowed Imaging schematic diagram of the elevation angle from -15 degree to 15 degree, Fig. 3 (b) represent imaging schematic diagram of the orientation from -150 meters to 150 meters, by Fig. 3 understands, is apparent from the point target that the inventive method shows.

It can be seen that, the requirement of forword-looking imaging can be met using Vivado HLS softwares in virtex7 Series FPGAs, it is to avoid The complex process of manual debugging Method at Register Transfer Level RTL, improves resource utilization, reduces time delay, greatly shorten exploitation Cycle.

Claims (3)

1. a kind of GRADIENT PROJECTION METHODS based on FPGA, including：

(1) obtain the echo data s of the target of actual measurement^uWith pattern data F_θ, wherein θ ∈ (1,10000), u ∈ (1,196)；

(2) calculate echo data s^u=F_θσ^uIn scene scatters factor sigma to be restored^u：

(2a) the first weighted factor β ∈ (0,1) of search and the second weighted factor are setSelect initial starting point z₀, And counter is set to 0；

(2b) calculate initial step length：a₀∈[α_min,α_max],

Wherein α_minRepresent and limit a₀The minimum of a value of scope, α_maxRepresent and limit a₀The maximum of scope, and 0<α_min<α_max, g_kIt is to throw The gradient direction of shadow,The transposition of T representing matrixs；

(2c) under conditions of the first weighted factor β and the second weighted factor μ is met, find the kth time that following formula is set up repeatedly Ride instead of walk long a_k：

$F\left({(z_k-a_k\&dtri;F\left(z_k\right))}_{+}\right)\&le;F\left(z_k\right)-\mathrm{\&mu;}\&dtri;F{\left(z_k\right)}^T(z_k-{(z_k-a_k\&dtri;F\left(z_k\right))}_{+})$

WhereinRepresent First-order Gradient value, ()₊Expression takes the positive portion of (), z_kRepresent feasible point during kth time iteration；

(2d) calculate new feasible pointz_k+1Represent feasible point during+1 iteration of kth；

(2e) convergence detection is carried out to feasible point, judges z_k+1Value whether meet following end condition：

WhereinIt is normal number, tolP is a minimum value parameter, if meeting the end Only condition, then export z_k+1Value, as scene scatters factor sigma to be restored^u；If be unsatisfactory for, k=k+1 is made, return to step Suddenly (2b)；

(3) step (2) is optimized in the Vivado HLS softwares that FPGA is carried, the scene to be restored after being optimized dissipates Penetrate coefficient

(4), after the completion of optimizing, Method at Register Transfer Level RTL is converted into into the knowledge with imaging function in Vivado HLS softwares Property right core, calls the IP core in FPGA, to optimization after scene scatters coefficient to be restoredRearranged simultaneously By vector superposed, the two dimensional image of scene to be restored is finally given, complete the gradient projection based on FPGA.

2. method according to claim 1, right in the Vivado HLS softwares that FPGA is carried wherein described in step (3) Step (2) is optimized, and carries out in accordance with the following steps：

(3a) loop unrolling instruction set_directive_unroll is selected in Vivado HLS softwares, by Do statement exhibition Open, realization completes circulate operation within the shorter clock cycle；

(3b) array split order array_partition is selected in Vivado HLS softwares, select in type Complete, selects 1 in dimension, and realization is loaded into all data within a clock cycle, improves operation efficiency；

(3c) instruction of data flow dataflow is selected in Vivado HLS softwares, realize executed in parallel, reach arithmetic speed It is most fast, finally give the scene scatters coefficient to be restored after optimization

3. method according to claim 1, calls IP core in FPGA wherein described in step (4), be The rising edge of clock signal sys_clk, when initial signal ap_start is high level, is input into the mesh of actual measurement to IP core Target echo data and pattern data, after one clock cycle, when idle signal ap_idle is low level, show to know Know property right core to work, as idle signal ap_idle and when completing signal ap_done and being high level, show intellectual property Core has completed work, the scene scatters coefficient to be restored after now the value of output data ap_return as optimizes