CN-121982127-A - Probability distribution sampling enhancement method based on photoelectric probability bit array in-situ sensing circuit

CN121982127ACN 121982127 ACN121982127 ACN 121982127ACN-121982127-A

Abstract

The invention provides a probability distribution sampling enhancement method based on a photoelectric probability bit array in-situ sensing circuit, which directly establishes a Sigmoid nonlinear mapping relation between incident light intensity and output probability on a device physical layer by utilizing photoelectric response characteristics and a random noise mechanism of photoelectric probability bit unit intrinsic of the circuit, and comprises optical image input, in-situ photoelectric response and probability bit stream generation, repeated continuous sampling and probability distribution integral processing on binary probability bit stream and image gray level reconstruction; the invention successfully recovers continuous analog gray information from random fluctuation noise flow which looks disorder with high fidelity, not only is an image enhancement means, but also is a bridge between probability calculation hardware and actual visual application, not only remarkably improves the recognition precision of the system in complex visual tasks, but also endows the system with strong image reconstruction capability and high-accuracy reverse image generation capability under the environment of serious original data loss.

Inventors

LIU FEI
CHEN WEICAN
XIA CHENHAO
FU YUNYI
ZHANG YUXIANG

Assignees

北京大学

Dates

Publication Date: 20260505
Application Date: 20251231

Claims (6)

1. A probability distribution sampling enhancement method based on a photoelectric probability bit array in-situ sensing circuit is characterized in that the in-situ sensing circuit comprises a reading circuit, a decoder, a grid voltage generator and a control module, wherein the reading circuit is composed of a photoelectric probability bit array, a comparator CMP and a reference voltage V ref , the photoelectric probability bit array is composed of a plurality of photoelectric probability bit units, the photoelectric probability bit units are used as core sensing components and comprise semiconductor devices with photoelectric response capability and random noise sources inside, the photoelectric probability bit units are provided with intrinsic Sigmoid photoelectric response characteristics and can convert continuously-changed analog incident light intensity into probability of an output state through a Sigmoid function, and the method comprises the following steps: S1, inputting an optical image, namely directly projecting a target optical image to be processed onto a photosensitive area of a photoelectric probability bit array, so that the light intensity distribution of the image is mapped into the physical state of each unit of the array in space; meanwhile, a control module is used for driving a grid voltage generator to apply bias voltage to each unit in the array, and the working points of each unit in the array are adjusted to a non-deterministic state sensitive to illumination, so that the random characteristic of a unit device is activated, the output state of the unit device can be randomly overturned under the driving of internal noise, and the statistical probability of overturning is directly modulated by the incident light intensity; S2, generating an in-situ photoelectric response and probability bit stream, namely utilizing the intrinsic photoelectric response characteristic of a photoelectric probability bit unit, wherein the photoelectric probability bit unit responds to the incident light intensity I in at the corresponding position input in the step S1, and the unit directly converts the incident light intensity into internal turnover probability P through the Sigmoid photoelectric response characteristic; S3, performing discrete sampling for a plurality of times, namely setting a time observation window containing N clock cycles, and performing continuous sampling for the probability bit stream generated in the step S2 for N times by using a comparator CMP and a reference voltage V ref under the drive of a control signal applied by a control module to obtain a binarized image sequence containing N Zhang Shunshi states; s4, probability distribution integration, namely, carrying out probability distribution integration processing on the N binarized image sequences acquired in the step S3 in a control module, specifically, carrying out time domain accumulation summation on logic state values, namely logic '0' or '1', which are output in N times of continuous sampling for each pixel point in the array, and dividing the accumulation result by the total sampling times N, so as to calculate and obtain time average probability P avg of the pixel point; S5, reconstructing a gray feature map, namely converting the time average probability P avg calculated in the step S4 into a corresponding pixel gray value according to a linear mapping rule, and reconstructing a target image which is identical to the original input image in vision.
2. The method of claim 1, wherein the optoelectric probability bit cell employs a voltage dividing structure of a series connection of a fixed resistor and a phototransistor, the fixed resistor being a stable load element, the phototransistor being exposed to external light for receiving an external light signal and modulating its own conductive characteristic according to light intensity.
3. The method of claim 1, wherein in the step S2, the decoder executes progressive scan logic driven by a control signal applied by the control module to sequentially activate the strobe signals of each row of the array, and performs progressive parallel reading and binarization on the analog potential of the output node of the photoelectric probability bit unit in cooperation with the comparator CMP and the reference voltage V ref to generate a random binary bit stream fluctuating with time, thereby realizing in-situ conversion from the optical signal to the probability bit stream.
4. The method of claim 1, wherein in the steps S1 and S2, the device adopted by the photoelectric probability bit unit has a Sigmoid-shaped light intensity-probability mapping relationship, specifically, when the incident light intensity I in is weak, the photoelectric probability bit unit is suppressed, the probability of outputting a logic "1" approaches 0, when the incident light intensity I in is strong, the probability of outputting a logic "1" approaches 1, when the incident light intensity is in the transition region, the probability of outputting a logic "1" and the light intensity show a Sigmoid-shaped rise, and the specific formula is as follows: Where P represents the probability that the cell outputs a particular logic state, I in is the incident light intensity, Representing the photo-coupling gain factor, which characterizes the sensitivity of the device to convert the optical signal into a probability change, determines the steepness of the activation function, Is a common system bias control quantity, and translates the position of the Sigmoid curve by adjusting the electrical bias.
5. The method according to claim 1, wherein in the steps S3 and S4, the validity of the strategy of N consecutive samples is based on the big theorem, and the reconstructed gray value G of the pixel is approximated by performing N consecutive samples and performing probability distribution integration: Wherein S i is Shan Xiangsu th sampling instant state, ES is the pixel output state expectation, P (I in ) is the output probability of photoelectric probability bit unit corresponding to current input light intensity I in , and the variance of measured value is increased along with the increase of sampling times N The formula shows that the finally reconstructed pixel gray value G is in a direct proportion relation with the theoretical probability P (I in ) corresponding to the current input light intensity, so that the original image gray information contained in the probability distribution is accurately read and restored.
6. The method according to claim 1, wherein in the step S5, the normalized probability value P avg is multiplied by the maximum gray value of the target image, so as to restore it to a standard digital gray value.

Description

Probability distribution sampling enhancement method based on photoelectric probability bit array in-situ sensing circuit Technical Field The invention belongs to the technical fields of artificial intelligent hardware, brain-like calculation and photoelectric information processing, and particularly relates to a probability distribution sampling enhancement method based on a photoelectric probability bit array in-situ sensing circuit. The method is particularly suitable for scenes such as edge calculation, intelligent sensing, machine vision, low-power consumption image reconstruction and recognition. Background With the proliferation of the number of nodes of the internet of things and the sinking of artificial intelligence applications, edge devices present serious challenges for high-energy-efficiency, low-delay sensing and computing capabilities. Probability computation (Probabilistic Computing) is used as an emerging paradigm for computation using physical randomness, by introducing probability bits (p-bits) as the basic computation units, using their random nature of fluctuation between 0 and 1 to model the activation behavior of neurons. The probability calculation shows the energy efficiency potential exceeding that of the traditional von neumann architecture on the tasks of solving Bayesian inference, combinatorial optimization (such as the solution of Xin Moxing) and generative model. However, existing probability computation hardware still presents significant bottlenecks in machine vision applications. Current probability bit devices (e.g., magnetic tunnel junctions, resistive random access memories, etc.) are mostly based on electrical mechanisms, lack of sensing capability, and must rely on external image sensors and analog-to-digital conversion circuits (ADCs). This "sense-store-calculate" physically separate architecture results in a large amount of data handling, introducing significant delay and energy consumption, namely von neumann bottlenecks. Meanwhile, a conventional photoelectric detector generally outputs deterministic signals, can not generate adjustable randomness and Sigmoid activation characteristics required by probability calculation, and is difficult to directly serve as input of probability neural networks such as a deep boltzmann machine. In addition, the probability bit output is essentially a randomly fluctuating signal, and the statistical characteristics of the probability bit output bear key information such as gray level, texture and the like of an image. However, the conventional binary sampling strategy can only acquire discrete states of single sampling, and cannot effectively capture continuous probability distribution information contained in random signals. For natural images containing complex gray scales and textures, the sampling mode can lead to the loss of key features, and severely limits the performance of a probability computing system in image reconstruction and recognition tasks. Therefore, a technical solution that can realize in-situ optical perception and preserve rich image features through an efficient probability distribution sampling strategy is needed to improve the performance of the probability computing system in visual intelligent application. Disclosure of Invention Aiming at the problems in the prior art, the invention provides a probability distribution sampling enhancement (Optical Probability Distribution Sampling Augmentation, abbreviated as OPDSA) method based on a photoelectric probability bit array in-situ sensing circuit. The method utilizes the intrinsic Sigmoid photoelectric response characteristic of the photoelectric probability bit device to realize in-situ sensing, and extracts image characteristics with high fidelity in the optical sensing process by a specific probability distribution sampling enhancement technology, so that the accuracy of image recognition and generation is improved. The invention can be realized by the following technical scheme: A probability distribution sampling enhancement (OPDSA) method based on an in-situ sensing circuit of a photoelectric probability bit array comprises a reading circuit, a decoder, a grid voltage generator and a control module, wherein the reading circuit is composed of the photoelectric probability bit array, a Comparator (CMP) and a reference voltage (V ref). The photoelectric probability bit array is composed of a plurality of photoelectric probability bit units, the photoelectric probability bit units are used as core sensing components, and the photoelectric probability bit array has intrinsic Sigmoid photoelectric response characteristics and can convert continuously-changed analog incident light intensity into probability of an output state through a Sigmoid function form. The OPDSA method flow is shown in figure 1 and comprises the following steps: S1, inputting an optical image, namely directly projecting a target optical image to be processed onto a photosensitive are