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

Abstract

The invention provides a probability distribution sampling enhancement method based on a photoelectric probability bit array in-situ sensing circuit, and the method comprises the steps: directly building a Sigmoid nonlinear mapping relation between incident light intensity and output probability at a physical level of a device through employing the intrinsic photoelectric response characteristics of a photoelectric probability bit unit of the circuit and a random noise mechanism; the method comprises the steps of optical image input, in-situ photoelectric response and probability bit stream generation, repeated continuous sampling and probability distribution integral processing on a binary probability bit stream, and image gray scale reconstruction. According to the method, continuous simulation gray information is successfully recovered from the noise flow which seems to be disordered and fluctuates randomly in a high-fidelity mode, the method is an image enhancement means and is a bridge between probability calculation hardware and actual visual application, the recognition precision of a system in a complex visual task is remarkably improved, and the recognition efficiency of the system is improved. And the system is endowed with strong image reconstruction capability and high-accuracy reverse image generation capability in the environment of serious loss of original data.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence hardware, neuromorphic computing, and optoelectronic information processing technology, specifically relating to a probability distribution sampling enhancement method based on an in-situ sensing circuit with an optoelectronic probability bit array. This invention is particularly applicable to scenarios such as edge computing, intelligent sensing, machine vision, and low-power image reconstruction and recognition. Background Technology

[0002] With the surge in the number of IoT nodes and the increasing adoption of artificial intelligence, edge devices face severe challenges in achieving high-efficiency, low-latency sensing and computing capabilities. Probabilistic computing, an emerging paradigm that leverages physical randomness, introduces p-bits as the basic computational unit, utilizing their random fluctuations between 0 and 1 to simulate neuron activation behavior. Probabilistic computing has demonstrated energy efficiency potential surpassing traditional von Neumann architectures in tasks such as Bayesian inference, combinatorial optimization (e.g., solving the Ising model), and generative models.

[0003] However, existing probabilistic computing hardware still faces significant bottlenecks in machine vision applications. Current probabilistic bit devices (such as magnetic tunnel junctions and resistive random access memories) are mostly based on electrical mechanisms and lack photosensitive capabilities, requiring external image sensors and analog-to-digital converters (ADCs). This physical separation of sensing, storage, and computation leads to a large amount of data transfer, introducing significant latency and energy consumption—the von Neumann bottleneck. Simultaneously, conventional photodetectors typically output deterministic signals and cannot intrinsically generate the tunable randomness and sigmoid activation characteristics required for probabilistic computation, making them unsuitable as direct inputs to probabilistic neural networks such as deep Boltzmann machines.

[0004] Furthermore, the output of probability bits is essentially a randomly fluctuating signal, and its statistical properties carry key information such as the image's grayscale and texture. However, traditional binarization sampling strategies can only acquire the discrete state of a single sample, failing to effectively capture the continuous probability distribution information inherent in random signals. For natural images containing complex grayscale and texture, this sampling method leads to the loss of key features, severely limiting the performance of probabilistic computing systems in image reconstruction and recognition tasks.

[0005] Therefore, there is an urgent need for a technical solution that can achieve in-situ optical sensing and retain rich image features through an efficient probability distribution sampling strategy, so as to improve the performance of probability computing systems in visual intelligence applications. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention proposes an Optical Probability Distribution Sampling Augmentation (OPDSA) method based on an in-situ sensing circuit using an optoelectronic probability bit array. This method utilizes the inherent sigmoid photoelectric response characteristics of the optoelectronic probability bit device to achieve in-situ sensing, and through specific probability distribution sampling augmentation techniques, extracts image features with high fidelity during optical sensing, thereby improving the accuracy of image recognition and generation.

[0007] This invention can be achieved through the following technical solutions: A probability distribution sampling enhancement (OPDSA) method based on an in-situ sensing circuit using a photoelectric probability bit array, wherein the in-situ sensing circuit includes a photoelectric probability bit array, a comparator (CMP), and a reference voltage (V). ref The system comprises a readout circuit, a decoder, a gate voltage generator, and a control module. The photoelectric probability bit array consists of multiple photoelectric probability bit units. These units, as the core sensing components, possess an intrinsic Sigmoid photoelectric response characteristic, enabling them to convert continuously varying simulated incident light intensity into the probability of the output state using a Sigmoid function. The OPDSA method flow is as follows: Figure 1 As shown, it includes the following steps: S1. Optical Image Input: The optical image of the target to be processed is directly projected onto the photosensitive area of ​​the photoelectric probability bit array, so that the light intensity distribution of the image is spatially mapped to the physical state of each unit in the array. Simultaneously, the control module drives the gate voltage generator to apply a bias voltage to each unit in the array, adjusting the operating point of each unit to a non-deterministic state sensitive to illumination. This activates the random characteristics of the unit devices, allowing their output states to randomly flip under internal noise, with the statistical probability of this flip being directly modulated by the incident light intensity.

[0008] S2. In-situ photoelectric response and probability bit stream generation: Utilizing the intrinsic photoelectric response characteristics of the photoelectric probability bit unit, the photoelectric probability bit unit responds to the incident light intensity at the corresponding position input in step S1. I in The unit directly converts incident light intensity into internal flip probability through the Sigmoid photoelectric response characteristic. P During this process, the decoder executes line-by-line scanning logic under the control signal applied by the control module, sequentially activating the gating signals of each row of the array; in conjunction with the comparator (CMP) and the reference voltage (V... refThe analog potentials of the output nodes of the photoelectric probability bit unit are read and binarized line by line in parallel. Without the need for an analog-to-digital converter (ADC), a random binary bit stream (a random sequence of 0s and 1s) that fluctuates over time is continuously and independently generated, achieving in-situ conversion from optical signals to probability bit streams. At this point, the overall output of the photoelectric probability bit array appears as a random binary image matrix modulated by light intensity. The logical state of each pixel is not fixed but flips between "0" and "1" with a specific statistical probability; the magnitude of this flipping probability directly reflects the light intensity at the corresponding location, thus realizing the output of a probability binary image.

[0009] S3. Multiple Discrete Sampling: A time observation window containing N clock cycles is defined. Driven by the control signal applied by the control module, a comparator (CMP) and a reference voltage (V) are used. ref The probability bitstream generated in step S2 is sampled N times consecutively to obtain a binary image sequence containing N instantaneous states. This step processes a discrete data sequence generated based on probability. Although the data sequence appears random in a single instance, its overall data set completely contains the light intensity feature information of the original image at a static moment. The more sampling times, the more complete the original image information it contains.

[0010] S4. Probability Distribution Integration: In the control module, probability distribution integration is performed on the N binarized image sequences obtained in step S3. Specifically, for each pixel in the array, the logical state values ​​(i.e., logic "0" or "1") output in N consecutive samples are summed in the time domain, and the summation result is divided by the total number of samples N to calculate the time-averaged probability of that pixel. P avg This operation essentially eliminates the quantization truncation error caused by a single sampling, thereby recovering the analog feature information that is highly correlated with the incident light intensity with high fidelity.

[0011] S5. Reconstruct the grayscale feature map: Reconstruct the time-averaged probability obtained in step S4. P avg Convert the values ​​to corresponding pixel grayscale values ​​according to the linear mapping rule. Specifically, the normalized probability values ​​are... P avg (Values ​​range from 0 to 1) Multiply by the maximum grayscale value of the target image to restore it to a standard digital grayscale value. This step reformats the processed probabilistic data into a common image data format, thereby reconstructing a target image that is visually consistent with the original input image. This reconstruction result can be used as high-quality deterministic input data, directly interfaced with backend digital image processing algorithms or probabilistic neural networks to support high-precision machine vision tasks.

[0012] Furthermore, in steps S1 and S2, the photoelectric probability bit unit employs a device with a sigmoid-shaped light intensity-probability mapping relationship. Specifically, when the incident light intensity... I in When the incident light intensity is low, the photoelectric probability bit unit is suppressed, and the probability of outputting logic "1" approaches 0; when the incident light intensity is low... I in When the incident light intensity is relatively strong, the photoelectric probability bit unit is activated, and the probability of outputting logic "1" approaches 1. When the incident light intensity is in the aforementioned transition region, the probability of outputting logic "1" increases in a Sigmoid manner with respect to the light intensity, as shown in the following formula: in P This represents the probability that the photoelectric probability bit unit outputs a logic "1". I in The incident light intensity is [value]. This represents the optocoupler gain coefficient, which characterizes the device's sensitivity in converting optical signals into probability variations and determines the steepness of the activation function. This is a general system bias control quantity. By adjusting the electrical bias, the position of the Sigmoid curve can be shifted. The above formula reflects the photo-probability response characteristics of the general photoelectric probability bit. By adjusting the output probability through the incident light intensity, the probability bit unit can sensitively distinguish the grayscale differences in the image and convert them into different probability distribution values.

[0013] Furthermore, in steps S3 and S4, the effectiveness of the N-times continuous sampling strategy is based on the Law of Large Numbers. If only conventional single-time sampling (i.e., N=1), i.e., the traditional binary sampling strategy, is used, the photoelectric probability bits can only output a discrete state of "0" or "1" at a certain moment. For image regions with rich gray-level gradations (such as shadow transitions) or fine textures, single truncation will lead to severe quantization noise, causing the intermediate gray-level information to be completely lost, and the image will present a rough black-and-white binary effect, which cannot meet the requirements of high-precision recognition.

[0014] By performing N consecutive samples and integrating the probability distribution, the reconstructed grayscale value of the pixel is obtained. G Approximate to: in, S i For a single pixel i The instantaneous state of the next sample. E [ S ] represents the desired output state of this pixel. P ( Iin () represents the current input light intensity I in The corresponding output probability of the photoelectric probability bit unit. This formula indicates the final reconstructed pixel grayscale value. G Theoretical probability corresponding to the current input light intensity P ( I in Proportional relationship This represents a direct proportional relationship, thus accurately reading and restoring the original image grayscale information contained in the probability distribution. As the number of samples N increases, the variance of the measured values... Gradually decrease.

[0015] The technical effects of this invention are as follows: This invention utilizes the inherent photoelectric response characteristics and random noise mechanism of photoelectric probability bit units to directly establish a sigmoid nonlinear mapping relationship between incident light intensity and output probability at the device physical level. This in-situ processing mechanism eliminates the need for independent optical sensors and high-energy-consuming analog-to-digital converters (ADCs) in traditional systems, thereby cutting off the massive data transfer path caused by the separation of sensing and computing at the source, effectively avoiding the von Neumann bottleneck, and significantly reducing the overall power consumption and hardware overhead of the system while eliminating data transmission delay. Based on this, the core of this invention introduces the Probability Distribution Sampling Enhancement (OPDSA) method, which is not only an image enhancement technique but also a bridge between probability computing hardware and practical vision applications. It creatively utilizes the probability distribution integral of the photoelectric probability bit output in the time domain to solve the fundamental contradiction that stochastic hardware cannot directly output high-fidelity deterministic images. By performing multiple consecutive samplings and probability distribution integration on the binary probability bitstream, the OPDSA method successfully recovers continuous analog grayscale information from a seemingly chaotic and randomly fluctuating noise stream. This effectively overcomes the quantization truncation error caused by traditional single-snapshot sampling, faithfully restoring and enhancing the fine texture features of the original image. This provides high-quality, detail-rich input data for backend probabilistic neural networks (such as deep Boltzmann machines). Benefiting from the dual advantages of hardware-level nonlinear activation and high-precision feature extraction, this invention not only significantly improves the system's recognition accuracy in complex visual tasks but also endows the system with powerful image reconstruction capabilities and high-accuracy inverse image generation capabilities even in environments with severely missing original data, greatly enhancing the robustness and intelligence of probabilistic computing vision systems. In summary, this invention, with its ultra-high energy efficiency and excellent feature processing capabilities, has broad application prospects in resource-constrained edge computing nodes, ultra-low-power intelligent IoT, unmanned autonomous systems, and next-generation neuromorphic vision chips. Attached Figure Description

[0016] Figure 1 This is a flowchart of the Probability Distribution Sampling Enhancement (OPDSA) method based on the in-situ sensing circuit of photoelectric probability bit array of the present invention. The OPDSA method includes steps such as multiple consecutive sampling, probability distribution integration operation and image grayscale reconstruction.

[0017] Figure 2 This is a schematic diagram of the in-situ optical sensing and probability mapping principle based on the in-situ sensing circuit of the photoelectric probability bit array of the present invention. It intuitively shows how the circuit uses the intrinsic sigmoid photoelectric response characteristics of the device to directly map the input continuous light intensity signal into a probabilistic binary data stream.

[0018] Figure 3 This is the output characteristic curve (optical Sigmoid curve) of the photoelectric probability bit unit (taking photoelectric IGZO TFT as an example) in this invention, which shows the Sigmoid-shaped mapping relationship of the probability of the output state being "1" as the incident light intensity changes.

[0019] Figure 4 These are comparison images showing the effect of image feature restoration using the OPDSA method of this invention. (a) is an unprocessed binary image of 8 single-sampled images, and (b) is the reconstructed image obtained after processing by the OPDSA method.

[0020] Figure 5 The graph shows the comparative test results of whether or not the OPDSA method of this invention is used on the accuracy of subsequent image recognition. The system using the OPDSA method achieved a high recognition rate of 91.66%, which is significantly better than the control group that did not use the method. Detailed Implementation

[0021] The present invention will be further illustrated below with examples. It should be noted that the purpose of disclosing the embodiments is to aid in further understanding the present invention; however, those skilled in the art will understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

[0022] Example 1: Verification of the intensity-probability mapping characteristics of photoelectric probability bits This example aims to illustrate how photoelectric probability bits can convert analog light intensity signals into a probability-tunable digital bit stream, and to verify that photoelectric probability bit units can utilize intrinsic randomness at the physical level to realize the Sigmoid mapping relationship between light intensity and probability distribution, which is the foundation for subsequent image processing.

[0023] The optoelectronic probability bit unit preferably employs a semiconductor device (e.g., a three-terminal field-effect device or a two-terminal resistive switching device) with photoelectric response capability and internal random noise sources. The incident light intensity modulates the device's random flip threshold or internal random noise sources. Physically, this unit is configured to operate in a nondeterministic state. When an analog light signal is incident, the intensity directly alters the statistical probability of the random signal within the unit crossing the decision threshold. This mechanism directly maps the amplitude of the light intensity in the analog domain to the statistical probability density of the logic state output in the time domain without analog-to-digital conversion (ADC). This example uses an optoelectronic probability bit unit based on an indium gallium zinc oxide (IGZO) thin-film transistor (TFT). The optoelectronic IGZO TFT device employs a bottom-gate top-contact structure, with an Al metal gate, an Al2O3 gate dielectric layer grown by ALD, an oxygen-rich vacancy-controlled IGZO thin film active layer, and Ti metal source and drain electrodes.

[0024] 1) Photoelectric probability bit unit setup: Construct photoelectric probability bit units composed of photoelectric IGZO TFTs. Each unit adopts a voltage divider structure (1T1R) with a resistor-transistor series connection. The upper part is connected to a fixed resistor as a stable load element, and the lower part of the photoelectric IGZO TFT is exposed to external light to receive external light signals and modulate its own conductivity according to the light intensity. The connection between the two is the intermediate node, which is also the output node of the photoelectric probability bit unit.

[0025] 2) Optical signal input and probability bit stream generation: A wavelength-tunable light source (e.g., 532nm) is used to illuminate the photoelectric probability bit unit with optical signals of different intensities (the light intensity range covers 0.1 to 1.1mW / cm²). 2 The optical signal directly acts on the photoelectric IGZO TFT, modulating its channel conductance (i.e., resistance value), while the series-connected fixed resistor maintains a stable resistance value. Together, they form a series voltage divider circuit. During this process, a bias voltage is applied to the gate of the photoelectric IGZO TFT using a gate voltage generator, modulating its operating point to a non-deterministic state sensitive to light, thereby activating the device's stochastic characteristics. At this time, the voltage divider value at the intermediate node of the cell is mainly determined by the light intensity and compared with a reference voltage. Because the device's operating point has been biased and adjusted to a non-deterministic state, the internal noise-driven stochastic characteristics are activated, resulting in weak random fluctuations superimposed on the voltage divider value at the intermediate node. This causes the output signal to exhibit a probabilistic bit stream that rapidly transitions between logic "0" and "1".

[0026] 3) Sigmoid mapping verification: Statistically determine the probability of the photoelectric probability bit unit outputting a "1" signal under 11 different light intensities (i.e., the proportion of times a "1" signal is output within a certain time period). Figure 3As shown, experimental measurements indicate that in the dark or under low light intensity (<0.2 mW / cm²), 2 At high light intensity (>1.0 mW / cm²), the probability of outputting "1" is close to 0%. As light intensity increases, the probability of outputting "1" increases in a sigmoid curve. 2 Under these conditions, the probability of outputting "1" approaches 100%. The goodness of fit (COD) of the Sigmoid curve for the experimental data exceeds 99%. This result confirms that the photoelectric probability bit unit can utilize the physical characteristics of the hardware itself to directly map a continuous light intensity signal into a nonlinear probability distribution signal.

[0027] Example 2: Image feature restoration effect based on OPDSA method This example aims to demonstrate how the Probability Distribution Sampling Enhancement (OPDSA) method can faithfully reconstruct the grayscale distribution and texture features of the original image from the discrete, random binary bitstream output by the photoelectric probability bit array through multiple sampling and probability distribution integration, and to verify its improvement on the accuracy of subsequent image recognition. The experiment is based on the photoelectric IGZOTFT-based photoelectric probability bit unit from Example 1, forming a 64×64 photoelectric probability bit array. The schematic diagram of the in-situ optical sensing and probability mapping based on the photoelectric probability bit array in-situ sensing circuit is shown below. Figure 2 As shown, the system drives and controls the optoelectronic probability bit array through collaborative addressing units (such as decoders), gate voltage generators, and peripheral readout circuits. Specifically, the decoder, acting as the row drive unit of the array, executes row-by-row scanning logic under the timing drive of control signals applied by the control module, sequentially activating the gating signals of each row of the array, thereby cooperating with the comparator (CMP) and reference voltage (V) of the top peripheral readout circuit. ref This system enables parallel, row-by-row reading of the potential of intermediate nodes in the photoelectric probability bit unit. Under the coordination of control signals, the above modules collaboratively complete the spatial projection of the optical signal, the probabilistic activation of the physical state of the photoelectric probability bit unit, and the quantization output of binary information. Finally, the analog potential suppressed by light intensity is quantized into a binary image matrix carrying probability distribution information, as shown on the right side of the figure. This example specifically includes: 1) Image Projection and Probability Bitstream Generation: A grayscale image of clothing (using an image from the Fashion-MNIST dataset, 64*64 pixels as an example) is projected onto the array's photosensitive surface through a lens. Simultaneously, a gate voltage generator, acting as the array's parameter configuration center, applies a configurable bias voltage to the gate of the photoelectric IGZO TFT in the photoelectric probability bit unit. This voltage sets the device's physical operating point by adjusting the channel conductance of the photoelectric IGZO TFT, and by adjusting the Sigmoid photoresponse curve, dynamic in-situ control of the array's sensing sensitivity and randomness intensity is achieved. Based on its intrinsic Sigmoid photoresponse characteristics, the photoelectric probability bit unit generates the incident light intensity... I in Directly mapped to internal flip probability P When the image at the corresponding location is brighter (i.e., the light intensity is higher), the more intense the light will be. I in The stronger the signal, the higher the statistical probability that the unit outputs a logic "1". P The higher the light intensity, the lower the probability of outputting logic "1"; conversely, the lower the light intensity, the lower the statistical probability of outputting logic "1". This rigorously transforms the spatial grayscale distribution of the original image into the probability distribution density of the output logic states of each pixel in the array. This is achieved in conjunction with a comparator (CMP) and a reference voltage (V). ref The analog potential of the intermediate node of the photoelectric probability bit unit is read and binarized line by line in parallel. A probability bit stream that fluctuates over time is continuously and independently generated without the need for an analog-to-digital converter (ADC).

[0028] 2) Binarization Sampling: If the system performs only a single sampling (N=1), that is, at a specific moment, the comparator is used to binarize the instantaneous fluctuations of the analog potential of each unit in the array. Since the device is in a non-deterministic state, this sampling forces the probability distribution state containing rich grayscale information to be compressed into a definite logic "0" or "1". The result is a binary image composed of black and white noise, such as... Figure 4 As shown in the single image in (a), the smooth grayscale gradations in the original image (such as the shadows of clothing folds) are completely lost, and the image appears as a coarse, grainy black and white image, making it difficult to distinguish the specific category of objects and texture details.

[0029] 3) Multiple Discrete Sampling and Probability Distribution Integration: A time observation window containing N clock cycles is set. Driven by the control signal applied by the control module, a comparator (CMP) and a reference voltage (V) are used. ref The output probability bitstream is sampled N times consecutively to obtain a sequence of N binarized images containing instantaneous states. In this embodiment, to visually demonstrate the effect, a sequence of N=8 consecutively acquired random binarized images (such as...) is selected. Figure 4(a) shows the process. A control module (such as a Field Programmable Gate Array (FPGA) or Microcontroller Unit (MCU) is used to integrate the probability distribution of these N images. Specifically, the binary data streams sampled multiple times are integrated and accumulated in the probability space to statistically determine the activation probability distribution of each pixel, and the time-averaged probability of each pixel in the array being in a specific logical state during N samples is calculated. P avg In this invention, the number of samples N is a configurable parameter. The system can dynamically adjust the value of N according to the actual application scenario: in scenarios with extremely high real-time requirements (such as high-speed motion capture), N is reduced to increase the frame rate; in scenarios with extremely high detail requirements (such as fingerprint recognition and medical imaging), N is increased to improve the signal-to-noise ratio. This dynamic reconfigurability makes this invention widely adaptable to various applications.

[0030] 4) Image reconstruction effect: The time-averaged probability calculated in the previous step is used to reconstruct the image. P avg Convert the values ​​to corresponding pixel grayscale values ​​according to the linear mapping rule. Specifically, the normalized probability values ​​are... P avg (Values ​​range from 0 to 1) Multiply by the maximum grayscale value of the target image (e.g., 255 for a commonly used 8-bit grayscale image), thus linearly quantizing it back to a standard digital grayscale value. The resulting reconstructed image exhibits rich grayscale levels, such as... Figure 4 As shown in (b), the random noise that was originally chaotic in a single frame image is smoothed and suppressed due to the statistical averaging effect; while the main outline, contrast and texture details of the object are clearly restored.

[0031] Impact of OPDSA on Image Recognition Results: To verify the actual contribution of the OPDSA method to the system's intelligent processing capabilities, this example sets up a comparative experiment based on the Fashion-MNIST dataset. The experiment constructed the same Deep Boltzmann Machine (DBM) network and divided it into two completely different training and testing procedures. The control group used single-sample (N=1) binarized images without OPDSA processing as the network's training and testing sets; the experimental group used reconstructed grayscale images enhanced by the OPDSA method (N=8) as the network's training and testing sets. The network contains a visible layer (1024 nodes) for receiving image data output from the photoelectric probability bit array, and two hidden layers (1504 nodes each) for feature extraction. The top layer also contains 10 label neurons for classification (corresponding to the 10 clothing categories in Fashion-MNIST).

[0032] Experimental results are as follows Figure 5As shown, the control group's recognition accuracy plateaued at approximately 86% due to the lack of features in its input data; while the experimental group, benefiting from the high-quality input data rich in grayscale details provided by OPDSA, was able to learn image features more effectively, ultimately achieving a high recognition accuracy of 91.66%. This significant performance improvement of approximately 5% powerfully demonstrates the core role of OPDSA technology in enhancing the robustness and accuracy of probabilistic computational vision systems.

[0033] This example fully demonstrates that although the instantaneous output of a single photoelectric probability bit is a binary random signal that has lost grayscale information, the system can effectively overcome the randomness limitations of the hardware layer by combining it with the OPDSA technology proposed in this invention. Through this strategy of multiple sampling and probability distribution integration, the system not only successfully extracts high-fidelity image grayscale and texture features from the probability distribution, but also significantly improves the recognition accuracy of the neural network trained and tested using this technology on the Fashion-MNIST dataset from approximately 86% in the control group to 91.66%. This result establishes the practical value and superiority of the probability distribution sampling enhancement method based on the in-situ sensing circuit of the photoelectric probability bit array described in this invention when performing high-precision, robust machine vision tasks.

[0034] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention, or modify them into equivalent embodiments, without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A probability distribution sampling enhancement method based on an in-situ sensing circuit with a photoelectric probability bit array, characterized in that, The in-situ sensing circuit includes a photoelectric probability bit array, a comparator CMP, and a reference voltage V. ref The method comprises a readout circuit, a decoder, a gate voltage generator, and a control module; wherein the photoelectric probability bit array is composed of multiple photoelectric probability bit units, and the photoelectric probability bit unit, as a core sensing component, includes a semiconductor device with photoelectric response capability and an internal random noise source; the photoelectric probability bit unit possesses an intrinsic Sigmoid photoelectric response characteristic, capable of converting continuously changing simulated incident light intensity into the probability of the output state through a Sigmoid function; the method includes the following steps: S1. Optical Image Input: The optical image of the target to be processed is directly projected onto the photosensitive area of ​​the photoelectric probability bit array, so that the light intensity distribution of the image is spatially mapped to the physical state of each unit of the array; at the same time, the control module drives the gate voltage generator to apply a bias voltage to each unit in the array, adjusting the operating point of each unit of the array to a non-deterministic state that is sensitive to light, thereby activating the random characteristics of the unit device, so that its output state can be randomly flipped under the drive of internal noise, and the statistical probability of this flipping is directly modulated by the incident light intensity; S2. In-situ photoelectric response and probability bit stream generation: Utilizing the intrinsic photoelectric response characteristics of the photoelectric probability bit unit, the photoelectric probability bit unit responds to the incident light intensity at the corresponding position input in step S1. I in The unit directly converts incident light intensity into internal flip probability through the Sigmoid photoelectric response characteristic. P At this time, the overall output of the photoelectric probability bit array is a random binary image matrix with light intensity modulation. The logical state of each pixel is not fixed, but flips between "0" and "1" with a specific statistical probability. The magnitude of the flipping probability directly reflects the light intensity at the corresponding position, thereby realizing the output of the probability binary image. S3. Multiple Discrete Sampling: A time observation window containing N clock cycles is set. Driven by the control signal applied by the control module, the comparator CMP and the reference voltage V are used. ref The probability bitstream generated in step S2 is sampled N times consecutively to obtain a binary image sequence containing N instantaneous states; S4. Probability Distribution Integration: In the control module, the probability distribution integration process is performed on the N binarized image sequences obtained in step S3. Specifically, for each pixel in the array, the logical state values ​​output in N consecutive samplings (i.e., logic "0" or "1") are summed in the time domain, and the summation result is divided by the total number of samplings N to calculate the time-averaged probability of that pixel. P avg ; S5. Reconstruct the grayscale feature map: Reconstruct the time-averaged probability obtained in step S4. P avg The target image is reconstructed by converting it into the corresponding pixel grayscale value according to the linear mapping rule, and then visually matching the original input image.

2. The method as described in claim 1, characterized in that, The photoelectric probability bit unit adopts a voltage divider structure of fixed resistor-phototransistor series connection. The fixed resistor serves as a stable load element, and the phototransistor is exposed to external light to receive external light signals and modulate its own conductivity according to the light intensity.

3. The method as described in claim 1, characterized in that, In step S2, the decoder executes line-by-line scanning logic under the control signal applied by the control module, sequentially activating the gating signals of each row of the array; in conjunction with the comparator CMP and the reference voltage V ref The analog potential of the output node of the photoelectric probability bit unit is read and binarized line by line in parallel to generate a random binary bit stream that fluctuates with time, thus realizing the in-situ conversion from optical signal to probability bit stream.

4. The method as described in claim 1, characterized in that, In steps S1 and S2, the photoelectric probability bit unit uses a device with a sigmoid-shaped light intensity-probability mapping relationship; specifically, when the incident light intensity... I in When the incident light intensity is low, the photoelectric probability bit unit is suppressed, and the probability of outputting logic "1" approaches 0; when the incident light intensity is low... I in When the incident light intensity is relatively strong, the photoelectric probability bit unit is activated, and the probability of outputting logic "1" approaches 1; when the incident light intensity is in the aforementioned transition region, the probability of outputting logic "1" increases in a Sigmoid manner with respect to the light intensity; the specific formula is as follows: in P This represents the probability that the unit outputs a specific logical state. I in For the incident light intensity, This represents the photoelectric coupling gain coefficient, a parameter that characterizes the device's sensitivity in converting optical signals into probability variations and determines the steepness of the activation function. It is a general system bias control quantity, which shifts the position of the Sigmoid curve by adjusting the electrical bias.

5. The method as described in claim 1, characterized in that, In steps S3 and S4, the effectiveness of the N-times continuous sampling strategy is based on the law of large numbers. By performing N-times continuous sampling and integrating the probability distribution, the reconstructed grayscale value of the pixel is obtained. G Approximate to: in, S i For a single pixel i The instantaneous state of the next sample. E [ S ] represents the desired output state of this pixel. P ( I in () represents the current input light intensity I in The output probability of the corresponding photoelectric probability bit unit; the variance of the measured value as the number of samplings N increases. Gradually decrease; this formula indicates that the final reconstructed pixel grayscale value G Theoretical probability corresponding to the current input light intensity P ( I in The probability distribution is directly proportional to the original image grayscale information, thus accurately reading and restoring the original image grayscale information contained in the probability distribution.

6. The method as described in claim 1, characterized in that, In step S5, the normalized probability value P avg Multiply by the maximum gray value of the target image to restore it to a standard digital gray value.

Images