Image reconstruction method of Poisson coding based on DHNN

Abstract

The invention relates to the technical field of Poisson coding, in particular to a DHNN-based Poisson coding image reconstruction method, which comprises the following steps: constructing a two-dimensional neuron DHNN model; generating two groups of hyper-chaos sequences by using a DHNN model, and inputting the two groups of hyper-chaos sequences into a pseudo-random number generator to obtain a group of pseudo-random numbers; comparing the pixel values of a plurality of pixel points of the grayscale image to be measured with the pseudo-random number to obtain a high-low level sequence of the corresponding pixel points; and reconstructing the grayscale image to be measured by using the high-low level sequence. According to the invention, a chaotic Poisson encoder is constructed by using a pseudo-random number generator based on a DHNN system, pulse signals are generated by using the Poisson encoder, and high and low levels are output for image reconstruction after the pulse signals are compared with pixel values of an image to be processed.

Description

Technical Field

[0001] The present invention relates to the technical field of Poisson coding, and in particular to an image reconstruction method of Poisson coding based on DHNN. Background Art

[0002] Spiking neural network (SNN), as the third generation of neural network, has attracted extensive research interest due to its significant advantages in resource cost and bionic capabilities; however, in the practical application of SNN, the encoding of the input signal is a crucial step. The encoder plays a key role in information transmission and storage. It converts raw data into a specific format in order to efficiently compress, transmit or store data.

[0003] The Poisson process is a statistical model that describes the number of random events that occur within a fixed time interval. Its main characteristics are the independence of events and the probability of an event occurring is proportional to the length of the time interval. Poisson coding is a coding method based on the Poisson distribution. As a stateless encoder, it is often used to preprocess the input signal of SNN. In Poisson coding, the probability of a pulse occurring conforms to the Poisson distribution, that is, within a given time interval, the number of pulses occurring follows the Poisson distribution law. The Poisson encoder converts the input signal into a peak rate λ (the number of pulses emitted within a given time T) and determines when to emit a spike by generating a pseudo-random number. In this way, the Poisson encoder can convert the input signal into a pulse signal for expression, which reduces the complexity of the input signal and improves the efficiency and effect of information processing and analysis.

[0004] Currently, most Poisson encoders for encoding SNN input signals are implemented using MCU and CPU. However, the encoders implemented using the above hardware platforms do not meet the expected levels in terms of power consumption and encoding time. For example, in Inherent Adversarial Robustness of Deep Spiking Neural Networks: Effects of Discrete Input Encoding and Non-Linear Activations, SNNs are difficult to train, which brings about a large amount of computing resource overhead. Summary of the invention

[0005] In view of the shortcomings of the existing methods, the present invention constructs a chaotic Poisson encoder based on the DHNN system using a pseudo-random number generator, uses the chaotic system to generate a highly random pseudo-random sequence, and outputs high and low levels after comparing it with the pixel values ​​of the image to be processed for image reconstruction.

[0006] The technical solution adopted by the present invention is: a DHNN-based Poisson coding image reconstruction method comprises the following steps:

[0007] Step 1: Construct a two-dimensional neuron DHNN model;

[0008] As a preferred embodiment of the present invention, the formula of the DHNN model is:

[0009]

[0010] Where μ is the decay of the neuron; a and b are the coupling connection weights; x n and n is the nth iteration value.

[0011] Step 2: Generate two sets of hyperchaotic sequences using the DHNN model, and input the two sets of hyperchaotic sequences into a pseudo-random number generator to obtain a set of pseudo-random numbers;

[0012] As a preferred embodiment of the present invention, the pseudo-random number generator is:

[0013]

[0014] Among them, K is the amplification factor; B is the bias; Indicates rounding down, and M is a constant.

[0015] Step 3: Compare the pixel values ​​of several pixels of the grayscale image to be tested with the pseudo-random number to obtain a high and low level sequence of the corresponding pixels;

[0016] As a preferred embodiment of the present invention, step three specifically includes:

[0017] Read the pixel value of the grayscale image pixel to be tested, compare the pixel value with the pseudo-random number, and generate a high level when the pixel value is less than the pseudo-random number, otherwise generate a low level.

[0018] As a preferred implementation of the present invention, the pixel value of the pixel point of the grayscale image to be measured is read at the PS end of the FPGA.

[0019] As a preferred implementation manner of the present invention, the pixel value is received at the PL end of the FPGA.

[0020] As a preferred embodiment of the present invention, the pixel value is sent from the PS end to the PL end via the axis-stream.

[0021] Step 4: Reconstruct the grayscale image to be tested using a high and low level sequence.

[0022] As a preferred embodiment of the present invention, a DHNN-based Poisson coding image reconstruction system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a DHNN-based Poisson coding image reconstruction method.

[0023] As a preferred embodiment of the present invention, a DHNN-based Poisson coding image reconstruction system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a DHNN-based Poisson coding image reconstruction method.

[0024] Beneficial effects of the present invention:

[0025] 1. The chaotic sequence generated by the discrete chaotic system has the characteristics of orbital boundedness, unpredictability and initial state sensitivity. It has passed the SP800-22 standard of the National Institute of Standards and Technology (NIST) of the United States. Compared with the traditional random sequence generated by computer, it has higher performance and is suitable for random sequence generation in Poisson encoders.

[0026] 2. The DHNN model is implemented using ZYNQ series ARM+FPGA heterogeneous hardware devices, which solves the limitations of traditional RTL-based FPGA development, such as low efficiency, cumbersome wiring, and single target application scenarios. Through efficient AXI bus connection, various application requirements can be flexibly realized;

[0027] 3. Realize the hardware of the chaos-based Poisson encoder, save the encoder requirements in the device in advance, and encode different input signals at any time, which greatly simplifies the complexity of the experiment. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of the image reconstruction method based on Poisson coding of DHNN of the present invention;

[0029] Figure 2 It is a dynamic behavior effect diagram of a two-dimensional discrete system generating a chaotic sequence according to the present invention;

[0030] Figure 3 The two-dimensional grayscale image of the present invention is an original image and a reconstructed image at different time nodes;

[0031] Figure 4 are the original image and the reconstructed image of the two-dimensional grayscale image of the present invention at different time steps;

[0032] Figure 5 It is the reconstructed image and pulse sequence captured by the hardware of the present invention in the oscilloscope;

[0033] Figure 6This is a comparison chart of the number of pulses observed in a certain period of time by Poisson coding of DHNN and the corresponding Poisson distribution fitting. DETAILED DESCRIPTION

[0034] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0035] like Figure 1 As shown, a DHNN-based Poisson coding image reconstruction method includes the following steps:

[0036] Step 1: Use activation functions sin and tanh to construct a DHNN model based on two-dimensional neurons;

[0037] The system equation of the two-dimensional discrete mapping model DHNN is as follows:

[0038]

[0039] Where μ is the decay of the neuron, μ∈(0,1); a and b are the coupling connection weights, specifying (a,b)=(10,1), (x 0 ,y 0 )=(1,1),x n and n is the nth iteration value.

[0040] Figure 2 The multi-cavity chaotic attractor diagrams with different numbers for different μ values ​​are given;

[0041] Step 2: The DHNN model provides two sets of hyperchaotic sequences X n and Y n , input two sets of hyperchaotic sequences into the pseudo-random number generator to obtain a set of pseudo-random numbers PRNs;

[0042] Definition {X n |n=1,2,3,…} and {Y n |n=1,2,3,…} are two sets of multi-fold hyperchaotic sequences provided by the DHNN model, namely X n By x n The sequence composed of n By n The sequence composed of n and Y n The two pseudo-random number generators (PRNG) can be designed as:

[0043]

[0044] Where K = 10 9is the amplification factor; B = 100 is the bias; It means rounding down. The range of pixel values ​​of the input two-dimensional grayscale image is [0, 255], so M = 256.

[0045] Due to its parallel computing capability, the FPGA hardware device based on the DHNN model can calculate the X-dimensional sequence and the Y-dimensional sequence in parallel and generate a set of PRNs at the same time. The hardware PRNs can be described as:

[0046]

[0047] When n=N, PRNs with a length of 2N can be obtained.

[0048] Step 3: Compare the pixel values ​​of several pixels of the grayscale image with the PRNs to obtain a high and low level sequence of the corresponding pixels;

[0049] The pixel value of the pixel point of the grayscale image is read at the PS end of the FPGA hardware device, and the pixel value is sent to the PL end through the axis-stream; when the PL end receives the pixel value of the grayscale image, the pixel value is compared with the pseudo-random number in the PRNs. When the pixel value is less than the pseudo-random number, a high-level signal is generated to the PS side, otherwise a low-level signal is generated; finally, the PS end outputs high-level and low-level signals through the DAC.

[0050] For example: PRNs = [200, 201, 220, 190, ..., 100], the pixel value of the first pixel of the grayscale image = 200, then the level sequence of the first pixel on the PS side is [0, 0, 0, 1, ..., 1].

[0051] Step 4: Use MatLab numerical simulation to obtain the original image and reconstructed image of the input two-dimensional grayscale image signal at different time nodes and step sizes.

[0052] like Figure 3 and Figure 4 As shown, the original image is used as a reference. Figure 3 (b) to (f) are the high and low level signal images reconstructed after encoding at different times, respectively. Figure 4(b) to (f) are reconstructed images after time step superposition from time 1 to 50; a two-dimensional grayscale image with a pixel size of 128×128 is selected as a test sample, in which the original image has complete signal information, while the image after Poisson encoding only contains black and white pixel information, that is, high and low level signals; different parts of the original image information are reconstructed at different time nodes, and then as the time step increases, the effect of the reconstructed image gradually improves and becomes closer and closer to the original image. This result is consistent with the characteristics of the traditional Poisson encoder, indicating that the chaos-based hardware Poisson encoder can effectively encode and reconstruct image information at different time steps.

[0053] The ZYNQ series ARM+FPGA heterogeneous hardware device is used to set up dual-channel output, obtain the pulse signal generated by the Poisson encoder, and display it on the oscilloscope.

[0054] Figure 5 The hardware platform and the image displayed on the oscilloscope are shown. Figure 5 (a) shows the reconstructed image captured by the oscilloscope in XY mode, corresponding to the numerical simulation part of step 3; the results verify the feasibility of the chaos-based hardware Poisson encoder; two AN9767 dual-channel DAC modules output time domain pulse trains for three different time nodes; Figure 5 (b) shows the time domain pulse sequences at t=1, 20 and 50, respectively, showing the time domain pulse sequences generated at different time nodes.

[0055] Figure 6 The figure shows the comparison of the number of pulses observed in a certain period of time and the corresponding Poisson distribution fitting. The time period T = 50 is selected (that is, the maximum number of pulses that can be observed is 50), and the horizontal axis represents the number of pulses that can be observed in all periods, ranging from 0 to 50; the vertical axis represents the number of pixels corresponding to the number of pulses generated in the input 128*128 image pixels, and the corresponding vertical axis data sum should be 128*128=16384. It can be seen from the figure that the number of pulses in a given time period follows the Poisson distribution law.

[0056] The performance comparison of the chaos-based Poisson encoder on different hardware devices is shown in Table 1:

[0057] Table 1 Performance of chaos-based Poisson encoder on different hardware devices

[0058]

[0059]

[0060] It can be seen that the computational power consumption of the Poisson encoder based on the ZYNQ series invention is comparable to that of the chaos-based hardware Poisson encoder implemented by a microcontroller (MCU), but its computational speed is tens or even thousands of times faster than that of the chaos-based hardware Poisson encoder implemented by a central processing unit (CPU) or MCU; this implementation not only proves the great potential of FPGA hardware devices in realizing SNN, but also further illustrates the superiority of this invention in terms of speed and power consumption.

[0061] The implementation of Poisson coding based on the ZYNQ platform not only benefits from the high programmability and parallel computing capabilities of the FPGA platform, but also because ARM can control the computing power. The combination of the two can significantly reduce power consumption and encoding time, further confirming the excellent performance of the chaos-based hardware Poisson encoder in practical applications.

[0062] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A DHNN-based Poisson coding image reconstruction method, characterized in that: The following steps are involved: Step 1: Construct a two-dimensional neuron DHNN model; Step 2: Generate two sets of hyperchaotic sequences using the DHNN model, and input the two sets of hyperchaotic sequences into a pseudo-random number generator to obtain a set of pseudo-random numbers; Step 3: Compare the pixel values ​​of several pixels of the grayscale image to be tested with the pseudo-random number to obtain a high and low level sequence of the corresponding pixels; Step 4: Reconstruct the grayscale image to be tested using a high and low level sequence.

2. The image reconstruction method based on Poisson coding of DHNN according to claim 1, characterized in that: The formula of the DHNN model is: Where μ is the decay of the neuron; a and b are the coupling connection weights; x n and n is the nth iteration value.

3. The image reconstruction method based on Poisson coding of DHNN according to claim 1, characterized in that: The pseudo-random number generator is: Among them, K is the amplification factor; B is the bias; Indicates rounding down, and M is a constant.

4. The image reconstruction method based on Poisson coding of DHNN according to claim 1, characterized in that: Step three specifically includes: Read the pixel value of the grayscale image pixel to be tested, compare the pixel value with the pseudo-random number, and generate a high level when the pixel value is less than the pseudo-random number, otherwise generate a low level.

5. The image reconstruction method based on Poisson coding of DHNN according to claim 4, characterized in that: The pixel value of the grayscale image pixel to be measured is read at the PS end of the FPGA.

6. The image reconstruction method based on Poisson coding of DHNN according to claim 5, characterized in that: Receive pixel values ​​on the PL side of the FPGA.

7. The image reconstruction method based on Poisson coding of DHNN according to claim 6, characterized in that: The pixel value is sent from the PS side to the PL side through the axis-stream.

8. The image reconstruction system based on Poisson coding of DHNN is characterized by: include: a memory for storing instructions executable by a processor; A processor, configured to execute instructions to implement the DHNN-based Poisson coding image reconstruction method as described in any one of claims 1-7.

9. A computer readable medium storing a computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements the image reconstruction method based on DHNN Poisson coding as described in any one of claims 1 to 7.

Images