A Fusion Method for Blurred Image Restoration and Recognition Based on Photon Reservoir Computation

Abstract

This invention relates to a fusion method for blurred image restoration and recognition based on photonic reservoir computing, comprising: S1: constructing an image restoration and recognition fusion model; S2: acquiring the blurred image to be processed; S3: in the image restoration stage, inputting the blurred image into the image restoration and recognition fusion model, and outputting a reconstructed clear image after processing through an input layer, a reservoir layer, and an output layer; S4: in the image recognition stage, feeding the reconstructed clear image back into the image restoration and recognition fusion model, and outputting the image recognition result after processing through an input layer, a reservoir layer, and an output layer. This invention can simultaneously realize blurred image restoration and recognition under a single photonic reservoir computing architecture. Compared with traditional separate architecture methods, it can complete the image deblurring and recognition fusion processing task at the device end with low power consumption and low latency, obtaining a reconstructed clear image and significantly improving the accuracy of image content recognition.

Description

Technical Field

[0001] This invention relates to the field of blurred image restoration and recognition technology, specifically to a blurred image restoration and recognition fusion method based on photon reservoir computing. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence, edge computing, and optoelectronic sensing technologies, intelligent recognition systems are being widely applied in complex scenarios such as autonomous driving, drone vision, satellite remote sensing imaging, and industrial inspection. Against this backdrop, real-time image processing and recognition capabilities in complex environments have become a core factor affecting system safety and reliability. In real-world scenarios, imaging systems are often affected by environmental factors such as high-speed motion, complex weather, low-light environments, scattering noise, and atmospheric disturbances, leading to severe blurring and degradation of acquired images and loss of detail, thus significantly reducing target recognition accuracy.

[0003] To address the aforementioned issues, traditional processing methods typically separate image restoration and image recognition into two relatively independent tasks, employing different system models or algorithm architectures to process them separately. However, this separate processing approach has significant limitations in applications such as autonomous driving, drone terminals, and industrial edge computing. First, the staged cascaded processing flow is complex. Under conditions of motion blur, low illumination, and high noise, the cascaded multi-module approach easily accumulates propagation errors, making it difficult to stably adapt to dynamic changes in real-world scenarios. Second, image restoration and target recognition models are often modeled and optimized independently. The former focuses on visual restoration quality, while the latter emphasizes target recognition accuracy. It is difficult to integrate and extract complementary information between image restoration and recognition, potentially resulting in enhanced image quality but limited improvement in subsequent image recognition accuracy. Finally, the separation of image restoration and recognition modules introduces additional storage, computing power, and data transmission overhead, increasing system latency and making it difficult to meet the requirements of device-side intelligent recognition systems for real-time performance, low power consumption, and lightweight implementation. Therefore, how to achieve fast, stable and highly accurate image deblurring and recognition fusion processing on low-energy-consumption and computing-limited terminal devices has become a core issue with important theoretical value and practical engineering significance in the field of intelligent recognition.

[0004] Traditional image deblurring methods can be broadly categorized into two types: one is based on statistical assumptions or specific prior models, such as methods based on maximum a posteriori probability (MAP) estimation, dark channel prior methods, and gradient-constrained optimization methods. These methods can effectively recover image structural information to some extent, but they typically rely on specific prior assumptions and their performance is limited in complex real-world scenarios. The other type is based on deep learning methods, such as convolutional neural networks (CNNs) and methods based on deep CNN denoisers. These methods have made significant progress in image deblurring tasks by learning mapping relationships in large-scale datasets. However, these methods usually rely on large amounts of high-quality training data, and the model training is complex and computationally expensive, so their application in real-world scenarios still faces challenges. In the field of image recognition, early methods mainly relied on design features and traditional machine learning models, such as scale-invariant feature transform (SIFT), histogram of oriented gradients (HOG), and support vector machines (SVM). These methods have simple structures and low computational costs, but their feature representation capabilities are limited, and their stability in recognizing scale changes, occlusion interference, and complex backgrounds is insufficient. In recent years, deep learning methods such as Convolutional Neural Networks (CNN), Residual Networks (ResNet), and VisionTransformer (ViT) have significantly improved image recognition performance through end-to-end feature learning, and have become the mainstream methods. However, deep learning networks typically have a large number of parameters, complex training processes, and high requirements for hardware resources and data scale, making their deployment in low-power, high-speed, real-time processing systems still challenging. Moreover, most current research still treats image deblurring and image recognition as two relatively independent tasks, failing to achieve the fusion of image deblurring and image recognition, which makes it difficult to meet the needs of current intelligent recognition systems for fast and high-precision deblurring and recognition of acquired images in real-world imaging environments.

[0005] The applicant discovered that Photonic Reservoir Computing (PRC), as an emerging photonic neural network computing method, has attracted widespread attention and made significant progress in fields such as time series prediction, target recognition, and nonlinear channel equalization due to its high-speed parallel processing capabilities and low training complexity. PRC systems typically construct high-dimensional dynamic systems based on optical devices such as semiconductor lasers, utilizing their inherent nonlinear response characteristics and short-time memory effects to achieve high-dimensional mapping and rapid processing of input signals.

[0006] Therefore, how to design a fuzzy image restoration and recognition fusion method based on photon reservoir computing is an urgent technical problem to be solved. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the technical problem to be solved by this invention is: how to provide a fusion method for fuzzy image restoration and recognition based on photon reservoir computing. This method can simultaneously realize fuzzy image restoration and recognition under a single photon reservoir computing architecture, making full use of the high-dimensional nonlinear mapping capability of photon reservoir computing. Compared with traditional separate architecture methods, this method can complete the fusion processing of image deblurring and recognition at the device end with low power consumption and low latency, resulting in a reconstructed clear image and significantly improving the accuracy of image content recognition.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] A method for blurry image restoration and recognition fusion based on photon reservoir computation includes:

[0010] S1: Construct an image restoration and recognition fusion model comprising an input layer, a reservoir layer, and an output layer; wherein, the input layer is used to perform pixel unpacking and masking on the input image to obtain the injection signal; the reservoir layer is used to perform nonlinear high-dimensional mapping on the injection signal of the input layer based on photon reservoir computation to obtain the state matrix; the output layer is used to generate the corresponding output result based on the state matrix of the reservoir layer and the trained output weights;

[0011] S2: Obtain the blurred image to be processed;

[0012] S3: In the image restoration stage, the blurred image to be processed is input into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding reconstructed clear image is output.

[0013] S4: In the image recognition stage, the reconstructed clear image is fed back into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding image recognition result is output.

[0014] Preferably, in step S1, the input layer expands the pixel values ​​corresponding to the input image into one-dimensional data, and then multiplies it with a preset six-value mask matrix to obtain the injected signal.

[0015] Preferably, in step S4, during the image recognition stage, the reconstructed clear image is first binarized, and then the binarized reconstructed clear image is input to the input layer for processing.

[0016] Preferably, in step S1, the reserve pool layer consists of several virtual nodes, and the virtual nodes are connected through randomly generated internal connection weights.

[0017] Each virtual node performs a nonlinear mapping on the input injection signal based on photon reservoir calculation to obtain the corresponding node response state;

[0018] The node response states of all virtual nodes together constitute the state matrix of the reserve pool layer and serve as the input to the output layer.

[0019] Preferably, the reservoir layer uses a discrete-mode semiconductor laser as the only physical node, and constructs multiple virtual nodes through a delay feedback mechanism. Each virtual node corresponds to the nonlinear dynamic response of the physical node at different time points, and together they form the high-dimensional state space of the reservoir layer.

[0020] Preferably, the processing steps of the reservoir layer include:

[0021] S301: Time The input signal is injected into the physical node for nonlinear high-dimensional mapping to obtain the corresponding node response state;

[0022] S302: After a fixed feedback time τ, the time will be... The node response status is fed back to the physical node via the feedback loop, and is synchronized with the time... The input signals work together on the physical node to perform a nonlinear high-dimensional mapping, resulting in a new node response state;

[0023] S303: Advance the moment to and return to step S302;

[0024] S304: Repeat steps S302 to S303 until all input signals have completed the corresponding nonlinear high-dimensional mapping and all node response states have been obtained;

[0025] S305: The state matrix of the reserve pool layer is composed of all the node response states and serves as the input to the output layer.

[0026] Preferably, the formula for nonlinear high-dimensional mapping of virtual nodes in the reserve pool layer is:

[0027] ;

[0028] ;

[0029] ;

[0030] In the formula: and These represent the slowly varying electric field and carrier density of a discrete-mode semiconductor laser, respectively. Linewidth enhancement factor; For injecting current; This is the gain coefficient; It represents the electron charge. Transparent carrier number; This is the gain saturation factor; Photon lifetime; This is due to frequency detuning; Injection intensity; For feedback strength; For feedback time; , , These are the nonradiative coefficient, the spontaneous coefficient, and the Auger recombination coefficient, respectively. The electric field strength that drives the laser; To drive the light intensity of the laser; For a moment Injection signal of the input layer.

[0031] Preferably, in step S1, the output layer calculates the state matrix output by the reservoir layer and the output weights obtained from training to obtain the corresponding output result, namely, the reconstructed clear image or the image recognition result.

[0032] During training, the output weights of the output layer of the image restoration and recognition fusion model are trained using the ridge regression calculation method.

[0033] Preferably, in step S1, when training the image restoration and recognition fusion model, peak signal-to-noise ratio, structural similarity index, and normalized mean square error are selected as evaluation indicators to assess its image restoration performance.

[0034] Preferably, in step S1, when training the image restoration and recognition fusion model, the recognition error rate is calculated using the 10-fold cross-validation method as an evaluation index to assess its image recognition performance.

[0035] Compared with existing technologies, the blurred image restoration and recognition fusion method based on photon reservoir computing in this invention has the following advantages:

[0036] This invention innovatively proposes a fuzzy image restoration and recognition method based on Photonic Reservoir Computing (PRC). Under a unified system architecture, it achieves rapid collaborative processing of fuzzy image restoration and recognition tasks, realizing the fusion of image deblurring and image recognition on the device side. On one hand, this invention employs a unified system architecture comprising an input layer, a reservoir layer, and an output layer, effectively reducing the computational overhead and processing latency introduced by the separate processing steps of traditional fuzzy image restoration and recognition models, while also improving the overall performance of image restoration and recognition. On the other hand, unlike existing PRC methods primarily applied to time-series signal processing or single image recognition tasks, this invention applies PRC to the fusion processing of fuzzy image restoration and recognition. Utilizing the high-dimensional nonlinear mapping capability of PRC, it enhances the image's recognizability while restoring the main structural information of the image, significantly improving the accuracy of subsequent image recognition tasks. Compared to mainstream deep learning methods, this invention does not require a large-scale and complex training process, nor does it rely on GPU computing platforms. It has advantages such as low computational cost and fast processing speed. It can meet the needs of fast, stable and high-precision device-side image processing and recognition under the constraints of low memory usage and limited computing power in intelligent recognition scenarios. It provides a new method with potential application value for fast image restoration and recognition in actual imaging environments. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0038] Figure 1 This is a logic block diagram of a fuzzy image restoration and recognition fusion method based on photon reservoir computation.

[0039] Figure 2 This is a diagram illustrating the change in PSNR with Node (the number below each image represents its corresponding PSNR).

[0040] Figure 3 This is a graph showing the trends of PSNR, SSIM, and NMSE as a function of Node.

[0041] Figure 4 This is a diagram illustrating how PSNR changes with the sample (the number below each image represents its corresponding PSNR).

[0042] Figure 5 This is a graph showing the trends of PSNR, SSIM, and NMSE as a function of Sample.

[0043] Figure 6 This is a schematic diagram of the experimental optical path.

[0044] Figure 7This diagram illustrates the variation of PSNR / SSIM / NMSE with Node and Sample (the number below each image represents its corresponding PSNR).

[0045] Figure 8 The following graphs show the trends of PSNR, SSIM, and NMSE with respect to NODE and SAMPLE: (a) shows the performance changes under different Node conditions; (b) shows the performance changes under different Sample conditions.

[0046] Figure 9 This is a comparison chart showing the effects of different methods on restoring blurred images.

[0047] Figure 10 This is a trend graph showing how WER changes with the binarization threshold.

[0048] Figure 11 Comparison of recognition results for blurred and restored images: (a) Recognition result of the unprocessed blurred image; (b) Recognition result after deblurring by the PRC system. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0050] The following detailed explanation illustrates the specific implementation methods:

[0051] Example:

[0052] This embodiment discloses a method for fuzzy image restoration and recognition based on photon reservoir computing.

[0053] like Figure 1 As shown, a method for blurry image restoration and recognition fusion based on Photonic Reservoir Computing (PRC) includes:

[0054] S1: Construct an image restoration and recognition fusion model (hereinafter referred to as the PRC system) containing an input layer, a reservoir layer, and an output layer. The input layer is used to perform pixel unpacking and masking on the input image to obtain the injection signal. The reservoir layer is used to perform nonlinear high-dimensional mapping on the injection signal of the input layer based on photon reservoir calculation to obtain the state matrix. The output layer is used to generate the corresponding output result based on the state matrix of the reservoir layer and the trained output weights.

[0055] S2: Obtain the blurred image to be processed;

[0056] S3: In the image restoration stage, the blurred image to be processed is input into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding reconstructed clear image is output.

[0057] S4: In the image recognition stage, the reconstructed clear image is fed back into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding image recognition result is output.

[0058] To better illustrate the technical solution of the present invention, this embodiment will be described in more detail through the following parts.

[0059] I. Input layer

[0060] In the specific implementation process, the input layer expands the pixel values ​​corresponding to the input image into one-dimensional data, and then multiplies it with a preset six-value mask matrix to obtain the injected signal.

[0061] In the image recognition stage, the reconstructed sharp image is first binarized, and then the binarized reconstructed sharp image is input into the input layer for processing. To simplify the data and facilitate the calculation of the storage pool, the reconstructed sharp image is binarized. After binarization, the pixel values ​​of the image are converted into a pixel matrix containing only 0s and 1s, which reduces data complexity and simplifies the calculation of the storage pool.

[0062] II. Reservoir layer

[0063] In the specific implementation process, the reservoir layer consists of several (non-linear) virtual nodes, which are connected by randomly generated internal connection weights. Each virtual node performs a non-linear mapping based on photon reservoir calculation on the input injection signal to obtain the corresponding node response state. The node response states of all virtual nodes together constitute the state matrix of the reservoir layer and serve as the input of the output layer.

[0064] The photonic reservoir layer uses a discrete-mode semiconductor laser as the sole physical node, and constructs multiple virtual nodes through a delay feedback mechanism. Each virtual node corresponds to the nonlinear dynamic response of the physical node at different time points, collectively forming the high-dimensional state space of the reservoir layer. In this single-physical-node photonic reservoir computation, the semiconductor laser, perturbed by optical feedback and injection, provides a nonlinear high-dimensional mapping for the reservoir layer. This high-dimensional mapping facilitates the rapid training of output connection weights, which are the only part that needs to be trained in this reservoir computation.

[0065] Specifically, the processing steps for the reservoir layer include:

[0066] S301: Time The input signal is injected into the physical node for nonlinear high-dimensional mapping to obtain the corresponding node response state;

[0067] S302: After a fixed feedback time τ, the time will be... The node response status is fed back to the physical node via the feedback loop, and is synchronized with the time... The input signals work together on the physical node to perform nonlinear high-dimensional mapping, resulting in the response state of the new node (the next virtual node);

[0068] S303: Advance the moment to and return to step S302;

[0069] S304: Repeat steps S302 to S303 until all input signals have completed the corresponding nonlinear high-dimensional mapping and all node response states have been obtained;

[0070] S305: The state matrix of the reserve pool layer is composed of all the node response states and serves as the input to the output layer.

[0071] Specifically, the formula for nonlinear high-dimensional mapping of virtual nodes in the reserve pool layer is as follows:

[0072] ;

[0073] ;

[0074] ;

[0075] In the formula: and These represent the slowly varying electric field and carrier density of a discrete-mode semiconductor laser, i.e., the response state. Linewidth enhancement factor; This refers to the injection current. This is the gain coefficient; It represents the electron charge. Transparent carrier number; This is the gain saturation factor; Photon lifetime; This is frequency detuning. Injection strength; For feedback strength; Feedback time; , , These are the nonradiative coefficient, the spontaneous coefficient, and the Auger recombination coefficient, respectively. The electric field strength driving the laser (Drive Laser, DL); To drive the light intensity of the laser; For a moment Injection signal of the input layer.

[0076] The fixed parameters are as follows: = 3, = 1.48 × 10 4 s -1 , = 1.93 × 10 7 , =7.73 × 10 -8 , = 2.17 ps, =1.6 × 10 -19 C, = 2.8 × 10 8 s -1 , = 9.8 s -1 , =3.84 × 10 -7 s -1 .

[0077] III. Output Layer

[0078] In the specific implementation process, the output layer calculates the state matrix output by the reservoir layer and the output weights obtained from training to obtain the corresponding output result, namely the reconstructed clear image or the image recognition result.

[0079] During training, the output weights of the output layer of the image restoration and recognition fusion model are trained using the ridge regression calculation method.

[0080] The ridge regression formula is expressed as:

[0081] (4)

[0082] In the formula: It is the target vector for training, that is, the target vector corresponding to the clear image; It is a training vector calculated by PRC, which is composed of vectors corresponding to blurred images through complex calculations; It is a small regularization constant, set here to 1.05 × 10⁻⁶. -3 ; It is an identity matrix.

[0083] IV. Evaluation Indicators

[0084] 1. For image restoration models:

[0085] To quantitatively evaluate the performance of the image restoration and recognition fusion model in blurred image restoration, this invention selects peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and normalized mean square error (NMSE) as evaluation metrics.

[0086] The PSNR is defined as follows:

[0087] (5)

[0088] in,

[0089] (6)

[0090] In the formula: Used to measure the quality of image restoration; This represents the mean square error between the original image and the restored image. and These represent the pixel positions of the original image and the restored image, respectively. grayscale value at that location and Indicates the size of the image. This represents the maximum number of pixels in the image, set to 255.

[0091] 2. For image recognition models:

[0092] When training the image restoration and recognition fusion model, the word error rate (WER) is calculated using the 10-fold cross-validation method as an evaluation metric for judging image recognition performance.

[0093] Identify the error rate The calculation formula is:

[0094] (7)

[0095] in, This indicates the number of incorrectly identified samples. This represents the total number of samples. The smaller the WER, the better the recognition performance of the image recognition model.

[0096] This invention uses handwritten digits as the input image to be recognized. To make the image recognition model more accurate and reliable, this invention employs an n-fold cross-validation method to test the system performance. n-fold cross-validation, also known as cyclic estimation, is a method in statistics for dividing a data sample set into smaller subsets. In this invention, n=10, and the handwritten digit recognition sample set is randomly and evenly divided into 10 subsamples using the 10-fold cross-validation method, numbered 1-10. The first subsample is selected as the test set, and the remaining subsamples (2-10) are used as the training set. The output weights are optimized using the training set, and the system's WER (Warnings Efficiency) is calculated using the test set, completing this system performance evaluation. The second subsample is used as the test set, and the remaining 9 subsamples are used as the training set for training, repeating this process 10 times. Each subsample is used as both a training set and a test set for validation, yielding an evaluation result each time. The average of the 10 evaluation results is taken as the final system performance evaluation result.

[0097] V. Experimental Instructions

[0098] This experiment mainly introduces the results of blurry image restoration and recognition tasks using the PRC system. The blurry image restoration task is divided into two parts: simulated data and experimental data.

[0099] 1. Blurred Image Restoration Task

[0100] The goal of this experiment is to input a blurred image within the established PRC system framework, process it through the PRC system, and output a restored image, that is, to perform a deblurring operation on a blurred image using the PRC system.

[0101] 1.1 Simulation Results

[0102] First, the MNIST handwritten digit dataset was used as the sample for the reconstruction task. MNIST is a handwritten digit database established by Google Labs and the Courant Institute at NYU, containing 70,000 handwritten digit images. The training set contains 60,000 images, and the test set contains 10,000 images, sourced from different schools and the Census Bureau. The clearest images from a portion of the MNIST handwritten digit dataset were directly selected for the training and test sets. Then, Gaussian blurring was applied to each clear image to obtain the blurred image training and test sets.

[0103] Figure 2This paper presents a comparison of the restoration results of typical handwritten digit images under different numbers of reservoir nodes. The left side shows the clear image (target image for restoration), the middle side shows the blurred image (input image), and the right side shows the restored image obtained through the PRC system (output image). It was found that as the number of nodes increases, the image outline gradually becomes clearer, and detailed information is significantly restored. This indicates that the PRC system's restoration performance for blurred digit images also improves. With 490 nodes, the PSNR of the restored image of the digit "7" reached 22.633. Furthermore, the restoration results for different digits show that the PRC system exhibits a consistent performance improvement trend across digits of varying shapes and structural complexities, demonstrating the method's good generalization ability.

[0104] To further explore the impact of node changes on the recovery performance of the PRC system, three evaluation metrics—PSNR, SSIM, and NMSE—were used to more comprehensively assess the recovery effect. Figure 3 The figure shows the trends of PSNR, SSIM, and NMSE as the number of nodes changes. With a fixed training set size, the changes in evaluation metrics were investigated as the number of virtual nodes increased from 10 to 900. Overall, the system performance exhibits a clear phased change characteristic with the increase in the number of nodes. When the number of nodes is small, PSNR rapidly increases from approximately 18 dB to around 20 dB, and SSIM also rises rapidly from 0.452 to around 0.759, while NMSE decreases significantly. This indicates that in a low-dimensional reservoir state, the system's nonlinear mapping capability is limited, making it difficult to fully characterize the complex information in the image degradation process. However, with the increase in the reservoir dimension, the system can capture more nonlinear features of the input signal, thus significantly improving the image restoration quality. When the number of nodes further increases to a medium scale, PSNR and SSIM continue to rise slowly, reaching peak values ​​of PSNR = 22.633 and SSIM = 0.844 at Node = 490, while NMSE drops to its lowest value of 0.081. This stage indicates that the high-dimensional mapping capability of the reservoir gradually approaches saturation, and the system reaches its optimal state in representing degraded information. At this point, the PRC system achieves the best results in restoring blurred images. However, as the number of nodes continues to increase to a higher scale, the changes in various evaluation indicators tend to level off, and even show slight fluctuations. This phenomenon suggests that when the dimension of the reservoir exceeds a certain range, the system performance improvement enters a saturation zone, and further increasing the number of nodes has limited effect on improving restoration performance. Considering practical factors such as restoration performance and computation time, Node = 490 is chosen as the optimal number of nodes for the PRC system. Furthermore, Figure 3The light-shaded areas around each curve represent error bands, providing a visual indication of the randomness or uncertainty associated with the five randomly generated mask signals. It can be seen that all three evaluation metrics achieved narrow error bands, demonstrating the stability of the PRC system in blurred image restoration tasks.

[0105] In addition, we also investigated the effect of the number of training samples on image restoration. Figure 4 This shows the recovery performance of the PRC system on typical handwritten digit images under different training sample numbers. When the number of training samples is small, although the PRC system can recover the basic outline of the digit, the detailed information of the recovered image is not complete. As the number of training samples increases, the digit edges become clearer, structural details are further recovered, and the quality of the recovered image is significantly improved. When Sample=500, the PSNR of the digit "7" reaches 23.552, and the recovered image is visually close to the original clear image. (Similar to...) Figure 2 The recovery results for different digit samples also show that the PRC system's performance in recovering blurred images improves with the increase in the number of training samples. Therefore, the PRC system exhibits a consistent performance improvement trend across digits of different shapes and complexities, indicating that the method has good generalization ability and stability.

[0106] To further investigate the impact of sample variations on the recovery performance of the PRC system, such as... Figure 5The figure shows the trend of PRC system performance indicators as a function of the number of training samples (Samples) under a fixed number of reserve pool nodes (Node=490). The overall trend shows that system performance significantly improves with increasing training sample count, then gradually stabilizes. When the number of samples is small, PSNR and SSIM are relatively low, while NMSE is high, indicating that PRC has not yet fully learned the mapping relationship between the input blurred image and the target sharp image. In this stage, due to insufficient training data, the output weight estimation has a large bias, resulting in unsatisfactory restoration results. As the number of samples increases to a moderate scale, PSNR and SSIM rapidly improve, and NMSE significantly decreases, indicating that the system performance enters a rapid convergence phase. This shows that with richer training data, the PRC system can more accurately learn the nonlinear mapping relationship in the image degradation imaging process, thereby significantly improving image restoration quality. As the number of training samples further increases, the changes in various performance indicators gradually level off: PSNR stabilizes in the range of approximately 22–23 dB, SSIM remains around 0.871, while NMSE tends to a lower, stable value. When the number of samples increases to 500, all evaluation indicators reach their optimal values ​​(PSNR = 23.552, SSIM = 0.913, NMSE = 0.056). This phenomenon indicates that under high sample size conditions, the system has essentially completed learning the input-output mapping relationship and has entered a performance saturation stage. Considering practical factors such as computation time, Sample = 500 was ultimately chosen as the training sample size for the PRC system.

[0107] 1.2 Experimental Results

[0108] After completing the numerical simulation based on the MNIST dataset, in order to further verify the applicability and stability of the proposed PRC system in a real-world environment, this experiment constructed an experimental optical path for acquiring experimental images and performing image reconstruction experiments. The optical path structure diagram is shown below. Figure 6As shown, the experimental system mainly consists of a laser source, a spatial filter, a beam expander, a spatial light modulator (SLM), a lens, and a CCD camera. The coherent light emitted by the laser first undergoes mode purification through the spatial filter to remove high-frequency noise and improve beam quality. Subsequently, the beam expander expands the beam, ensuring it uniformly illuminates the SLM. The SLM loads the target image and modulates the random scattering mode; its output light field propagates through the lens system and forms an image on the CCD camera plane. After acquisition, the experimental image acquired by the CCD camera is input into the photon reservoir computing system, which uses its high-dimensional nonlinear mapping capability to reconstruct a clear image from a blurred one. This experimental platform simulates a real imaging environment, providing a reliable foundation for verifying the effectiveness of the PRC system in actual imaging environments. Compared to the numerical simulations described earlier, the actual optical path inevitably introduces environmental noise, device errors, and system instability, which will have a certain impact on image restoration performance. Therefore, the results obtained based on this experimental platform can more realistically reflect the performance of the proposed method in real-world applications.

[0109] The experimental optical path was constructed to acquire actual images, which were then input into the PRC system for reconstruction. Figure 7This paper presents the restoration results of typical digital images and their corresponding PSNR metrics under different numbers of nodes and training samples. The overall results show that the PRC system can still effectively restore image structure under the experimental data conditions. With a fixed number of samples, the restored image quality significantly improves as the number of nodes increases from 10 to 600. When the number of nodes is 10, the restored result still suffers from strong noise interference, and the digital structure is relatively blurry, with a corresponding PSNR of 19.382 dB. When the number of nodes increases to 100, the contours of the restored image gradually become clearer, and the PSNR significantly improves to 23.479 dB. When the number of nodes further increases to 600, image details are further restored, and the PSNR reaches 23.639 dB. This is because increasing the size of the storage pool helps improve the system's ability to express complex degradation features. Under the same number of nodes, increasing the number of samples also effectively improves the quality of the restored image. With 600 nodes, the PSNR improved from 23.639 dB at Sample=10 to 24.973 dB at Sample=20, and further to 25.515 dB at Sample=100. Visually, as the number of training samples increases, background noise gradually decreases and structural information becomes more complete. Results from different node and sample combinations reveal a synergistic effect on system performance. This is because larger nodes enable the PRC system to have stronger high-dimensional nonlinear mapping capabilities, while larger samples help the PRC system learn more fully the complete mapping relationships, improving generalization ability and the accuracy of output weight training, thus jointly enhancing the recovery performance of the PRC system.

[0110] To further investigate the impact of the number of nodes in the storage pool and the number of training samples on the recovery performance of the PRC system under experimental data conditions, such as Figure 8 The figure shows the trend of PRC system performance evaluation indicators (PSNR, SSIM, and NMSE) as a function of Node and Sample. Figure 8 In the diagram (a), the performance variation is represented under different Node conditions. Figure 8 (b) represents the performance variation under different sample conditions. Figure 8 As can be observed in Figure A, NMSE decreases rapidly with the increase of Nodes, while PSNR and SSIM increase significantly, reaching their optimal performance around Node = 600. This indicates that increasing the number of nodes in the reserve pool can effectively improve the system's high-dimensional nonlinear mapping capability. However, when the number of Nodes increases further, the performance indicators tend to level off, indicating that the system gradually enters the performance saturation region. Figure 8Figure (b) illustrates the performance trend of the system under different training sample sizes. When the sample size is small, the system performance improves rapidly with increasing training sample size, as evidenced by a significant increase in PSNR and SSIM and a rapid decrease in NMSE. However, when the sample size exceeds 100, the performance improvement slows down, with slight fluctuations in the high-sample region. This indicates that under the current experimental conditions, the system is able to learn the main mapping relationships with a small sample size, and further increasing the data volume has limited contribution to performance improvement. In summary, Figure 8 The results show that, under the experimental data conditions, both Nodes and Samples significantly impact system performance, but in different ways. Nodes primarily affect the model's expressive power, while Samples affect the sufficiency of model training. Notably, there are discrepancies between experimental and simulated results when searching for the optimal values ​​of the number of nodes in the buffer pool and the number of training samples. This is because factors such as multiple scattering, coherent noise, device errors, and environmental disturbances in the experimental optical path lead to more severe degradation of the input image, reducing effective information and thus making it more reliant on increasing the number of nodes in the buffer pool to enhance high-dimensional mapping capabilities. Simultaneously, increasing the number of samples in a high-noise environment provides limited information gain and easily introduces more noise disturbances. Therefore, increasing the number of training samples does not necessarily lead to performance improvement; instead, it may cause the system to overfit noisy features, resulting in performance fluctuations and affecting the stability of ridge regression weight estimation.

[0111] To further verify the effectiveness of this method, this study selected several typical deblurring methods for comparison, including Bayesian MAP estimation, dark channel prior, deep CNN denoising prior, and optimization methods based on cross-partial derivatives. Figure 9 As shown, different deblurring methods are applied to the same experimental dataset (by...). Figure 6The comparison of deblurring effects on the experimental optical path acquisition is shown in Table 1. Table 1 presents the specific evaluation metrics for the restoration effect of each method on the blurred image. As can be seen from the table, in terms of PSNR, PRC reaches 25.515 dB, an improvement of more than 4 dB compared to the better-performing deep CNN denoising method (21.478 dB), indicating its significant advantage in restoring image details and suppressing noise. In terms of SSIM, PRC reaches 0.700, significantly higher than other methods, indicating that the restored image is closer to the original image in terms of structure preservation. Meanwhile, in terms of NMSE, PRC is only 0.096, significantly lower than other methods, further verifying its high reconstruction accuracy. It is worth noting that traditional methods (such as Bayesian MAP estimation) perform relatively poorly on this experimental data. This is because traditional deblurring methods rely on specific statistical assumptions and prior models, but these assumptions are often difficult to hold under actual imaging conditions. In contrast, although the deep CNN denoising method improves performance to some extent, it is still limited by the distribution of training data and model structure, and its adaptability to high-noise degraded data is limited. Compared to the methods mentioned above, the PRC system transforms the complex degradation process into a linearly separable problem through high-dimensional nonlinear dynamic mapping and combines it with ridge regression to achieve efficient training. Therefore, it is better able to adapt to the characteristics of experimental data, such as strong scattering, high noise, and multiple environmental perturbations. Furthermore, PRC only needs to train the output weights, avoiding the large-scale parameter optimization process of deep learning networks, and it does not require the use of GPUs for large-scale training, significantly reducing computational costs and time. It has significant advantages in both computational complexity and training efficiency.

[0112] Table 1 Comparison of blurry image restoration effects with other methods

[0113]

[0114] 2. Recover the image recognition task

[0115] Building upon the previously presented task of restoring blurred images, this experiment further evaluates the effectiveness of blurred image restoration in practical applications. It performs image recognition tasks on both the collected blurred handwritten digit images and those deblurred using the PRC system. In image recognition, the representation of the input data significantly impacts the final recognition performance. The PRC system learns the nonlinear relationship between input and output through high-dimensional dynamic mapping. When the input data contains a large amount of noise or irrelevant grayscale variations, it reduces the discriminative power of the buffer pool. Binarization effectively enhances the target's structural information, making the target contour clearer and improving the separability of image features. Binarization also reduces the complexity of the input data; the effective information in the input dimension is compressed, helping to reduce the impact of redundant information on model training, thereby improving the system's efficiency in learning mapping relationships and its recognition performance.

[0116] Figure 10 The figure shows the trend of WER (Warning Error Rate) of the PRC system with respect to the binarization threshold under experimental data conditions. As can be seen from the figure, WER exhibits a clear trend of first decreasing and then increasing with the threshold. When the threshold is low, WER is at a high level, indicating poor recognition performance. As the threshold gradually increases, WER continuously decreases, reaching its minimum value at a threshold of 170, corresponding to optimal recognition performance. When the threshold is further increased, WER rises again, indicating that recognition performance begins to decline. This shows that the binarization threshold has a significant impact on the recognition results of the PRC system, and both excessively high and excessively low thresholds negatively affect recognition performance. When the threshold is too low, a large amount of background noise is misidentified as image structure, leading to an increased recognition error rate; conversely, when the threshold is too high, some effective structures are removed, resulting in missing structural information, which also reduces recognition accuracy. Therefore, there exists an optimal binarization threshold that minimizes the WER of the PRC system and achieves the best recognition performance. It is worth noting that the optimal binarization threshold varies significantly under different data conditions. For the experimentally acquired images, the optimal threshold is approximately 170, while for the MNIST database recognition task, the optimal threshold is approximately 140. This is because the experimentally acquired images have a higher degree of blur and noise level compared to the MNIST simulated data. Due to multiple scattering, coherent noise and systematic errors in the actual imaging process, the overall grayscale distribution of the image is lower and the contrast is reduced. In this case, a higher binarization threshold is needed to effectively suppress background noise and thus highlight the target structure.

[0117] like Figure 11 The figure shows a comparison of the recognition confusion matrices of the blurred images acquired in the experiment before and after deblurring. In the figure, (a) represents the recognition result of the unprocessed blurred image, and (b) represents the recognition result after deblurring using the PRC system. Figure 10As shown in the confusion matrix (a), under blurred image conditions, there is a significant misclassification phenomenon among the categories. The off-diagonal elements in the matrix are widely distributed, indicating severe confusion in the recognition results. In contrast, after deblurring by the PRC system, the confusion matrix exhibits a significant diagonal dominance, with the vast majority of samples being correctly classified and the off-diagonal elements significantly reduced. This result demonstrates that the PRC system can effectively recover key structural features in blurred images, improving the accuracy of subsequent recognition. To further quantitatively analyze the improvement in recognition performance by the PRC system, the WER of handwritten digit recognition was calculated to specifically evaluate the system performance, as shown in Table 2. The table shows that the WER of blurred images is only 0.672, indicating that the input blurred images are difficult to classify effectively. After deblurring by the PRC system, the WER of the recovered image significantly decreased to 0.134, a reduction of 80.1% compared to the WER before deblurring. This result demonstrates that the PRC system can effectively recover key structural information in images, significantly improving the separability of input data, thereby significantly improving recognition accuracy. The results in Table 2 verify the effectiveness of the proposed method. The PRC system can not only achieve high-quality image restoration, but also significantly improve the accuracy of subsequent recognition tasks, demonstrating its potential application value in real imaging environments.

[0118] Table 2. Error Rate of Handwritten Digit Recognition Using Blurred Images and PRC Reconstructed Images

[0119]

[0120] This invention proposes an end-to-end fusion processing method for blurred image restoration and recognition based on photon reservoir computation using semiconductor lasers. Simulated data and experimentally acquired data are used to verify the effectiveness and application potential of the PRC system in the field of image processing. First, this invention constructs a simulated dataset based on the MNIST dataset and systematically analyzes the impact of the number of virtual nodes in the reservoir and the number of training samples on system performance, verifying the feasibility of the PRC system in image deblurring tasks. Results show that as the number of nodes and samples increases, the overall restoration performance of the system improves, reaching its optimal state when Node=490 and Sample=500. To further verify the applicability of the PRC system in real-world scenarios, this invention constructs an experimental optical path and acquires actual degraded images for research. Results show that under the experimental data conditions, the optimal parameters for the system are Node=600 and Sample=100, at which point the PSNR of the restored image reaches its maximum value of 25.515 dB. By comparing the optimal parameter values ​​of simulated and experimental data, it was found that when the input image is subjected to stronger blurring and noise interference, increasing the pool dimension can more effectively enhance the system's ability to express complex nonlinear mappings, while the benefit of simply increasing the number of training samples tends to be limited. In subsequent recognition tasks, experimental data was used as the dataset to study the impact of the binarization threshold on the system's recognition performance. It was found that the best recognition effect was obtained when the threshold was 170. Compared with the unprocessed blurred image, the image recognition performance after PRC restoration was significantly improved, with WER significantly reduced from 0.672 to 0.134.

[0121] In summary, the method proposed in this invention can achieve enhanced performance in blurry image restoration and recognition under actual imaging conditions, and exhibits good generalization ability and stability, proving that the PRC system has potential application value in the field of image processing.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A blurred image restoration and recognition fusion method based on photonic reservoir computing, characterized in that, include: S1: Construct an image restoration and recognition fusion model comprising an input layer, a reservoir layer, and an output layer; wherein, the input layer is used to perform pixel unpacking and masking on the input image to obtain the injection signal; the reservoir layer is used to perform nonlinear high-dimensional mapping on the injection signal of the input layer based on photon reservoir computation to obtain the state matrix; the output layer is used to generate the corresponding output result based on the state matrix of the reservoir layer and the trained output weights; S2: Obtain the blurred image to be processed; S3: In the image restoration stage, the blurred image to be processed is input into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding reconstructed clear image is output. S4: In the image recognition stage, the reconstructed clear image is fed back into the image restoration and recognition fusion model. After processing by the input layer, the reservoir layer and the output layer, the corresponding image recognition result is output.

2. The photonically based reservoir computing method for fused image recovery and recognition of claim 1, wherein: In step S1, the input layer expands the pixel values ​​corresponding to the input image into one-dimensional data, and then multiplies it with a preset six-value mask matrix to obtain the injected signal. 3.The method of claim 1, wherein the method further comprises: determining a plurality of candidate images based on the blurred image; and determining a final image based on the plurality of candidate images. In step S4, during the image recognition stage, the reconstructed clear image is first binarized, and then the binarized reconstructed clear image is input into the input layer for processing.

4. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 1, wherein: In step S1, the reserve pool layer consists of several virtual nodes, and the virtual nodes are connected through randomly generated internal connection weights. Each virtual node performs a nonlinear mapping on the input injection signal based on photon reservoir calculation to obtain the corresponding node response state; The node response states of all virtual nodes together constitute the state matrix of the reserve pool layer and serve as the input to the output layer.

5. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 4, wherein: The reservoir layer uses a discrete-mode semiconductor laser as the only physical node and constructs multiple virtual nodes through a delay feedback mechanism. Each virtual node corresponds to the nonlinear dynamic response of the physical node at different time points, and together they form the high-dimensional state space of the reservoir layer.

6. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 5, wherein: The processing steps for the reservoir layer include: S301: Time The input signal is injected into the physical node for nonlinear high-dimensional mapping to obtain the corresponding node response state; S302: After a fixed feedback time τ, the time will be... The node response status is fed back to the physical node via the feedback loop, and is synchronized with the time... The input signals work together on the physical node to perform a nonlinear high-dimensional mapping, resulting in a new node response state; S303: let the time advance to and return to step S302; S304: Repeat steps S302 to S303 until all input signals have completed the corresponding nonlinear high-dimensional mapping and all node response states have been obtained; S305: The state matrix of the reserve pool layer is composed of all the node response states and serves as the input to the output layer.

7. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 6, wherein: The formula for nonlinear high-dimensional mapping of virtual nodes in the reserve pool layer is: ； ； ； In the formula: and These represent the slowly varying electric field and carrier density of a discrete-mode semiconductor laser, respectively. Linewidth enhancement factor; For injecting current; This is the gain coefficient; It represents the electron charge. Transparent carrier number; This is the gain saturation factor; Photon lifetime; This is due to frequency detuning; Injection intensity; For feedback strength; For feedback time; , , These are the nonradiative coefficient, the spontaneous coefficient, and the Auger recombination coefficient, respectively. The electric field strength that drives the laser; To drive the light intensity of the laser; For a moment Injection signal of the input layer.

8. The method for fuzzy image restoration and recognition based on photon reservoir computing as described in claim 1, characterized in that: In step S1, the output layer calculates the state matrix output by the reservoir layer and the output weights obtained from training to obtain the corresponding output result, namely the reconstructed clear image or the image recognition result. During training, the output weights of the output layer of the image restoration and recognition fusion model are trained using the ridge regression calculation method.

9. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 1, wherein: In step S1, when training the image restoration and recognition fusion model, peak signal-to-noise ratio, structural similarity index, and normalized mean square error are selected as evaluation indicators to assess its image restoration performance.

10. The photonically based reservoir pool computing method for fused blur image restoration and recognition of claim 1, wherein: In step S1, when training the image restoration and recognition fusion model, the recognition error rate is calculated using the 10-fold cross-validation method as an evaluation metric for assessing its image recognition performance.

Images