An image recognition method and system based on an in-vitro neuronal network reservoir computing architecture
Patent Information
- Application Number
- CN202610727279.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-25
- Publication Date
- 2026-08-18
AI Technical Summary
[0004]然而，现有基于体外神经元网络的储备池计算多集中于一维时序信号处理
[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description.
Smart Images
Figure CN122598167A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biological reservoir computing technology, and in particular to an image recognition method and system based on an in vitro neural network reservoir computing architecture. Background Technology
[0002] With the development of technologies such as image recognition, speech processing, and natural language understanding, artificial intelligence systems are placing higher demands on computing power, storage resources, and energy consumption control. Traditional artificial neural networks typically rely on large-scale matrix operations and continuous training, and in embedded intelligent sensing, low-power terminals, and bio-compatible computing scenarios, they still face problems such as high energy consumption, high training costs, and complex hardware deployment.
[0003] In vitro cultured biological neural networks possess self-organizing connections, nonlinear dynamics, short-term memory, and high parallel information processing capabilities, making them a viable physical implementation of reservoir computing. The basic idea of reservoir computing is to map input signals to a high-dimensional state space using a reservoir with complex dynamic characteristics, and then perform classification or prediction through a relatively simple output layer. Therefore, in vitro neural networks and reservoir computing have a high structural fit.
[0004] However, existing reservoir computing based on in vitro neural networks mostly focuses on one-dimensional temporal signal processing. For two-dimensional image recognition tasks, there is still a lack of clear and feasible processing chains for how to convert image spatial features into spatiotemporal stimulus signals that can be stably responded to by neural networks, and how to read the true dynamic response of neural networks while avoiding stimulus artifacts. Summary of the Invention
[0005] In view of this, embodiments of the present invention provide an image recognition method based on an in vitro neural network reservoir computing architecture to eliminate or improve one or more defects existing in the prior art.
[0006] One aspect of the present invention provides an image recognition method based on an in vitro neural network reservoir computing architecture, the method comprising the following steps: Obtain the image sample to be identified, and perform size normalization and grayscale normalization on the image sample to be identified to obtain the image grayscale matrix; The image grayscale matrix is divided into multiple region vectors, and the stimulation order of the region vectors is determined based on the structural order of the image grayscale matrix. The inner radius and width of the corresponding annular light spot are determined based on the gray value of each pixel in the region vector. The annular light spots of the region vector are combined to obtain the stimulation pattern of the region vector. Based on the stimulation sequence, light stimulation is applied to a preset in vitro neuronal network reserve pool layer according to a preset column stimulation cycle, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern. The calcium fluorescence response generated by the in vitro neuronal network reserve pool layer is acquired by a calcium imaging readout module equipped with a camera, and constructed into a calcium fluorescence response sequence. Based on the calcium fluorescence response sequence of each candidate neuron, a relative fluorescence change sequence is determined, and the relative fluorescence change sequence is input into a preset classification model. The classification model then outputs the classification result.
[0007] Using the above scheme, the image sample to be identified is first converted into an image grayscale matrix. Then, the image grayscale matrix is converted into a spatiotemporal sequence of light stimulation patterns in the form of ring-shaped light spots according to the structural order. This allows the spatial grayscale changes of the two-dimensional image to enter the in vitro neural network reserve pool layer in both temporal order and spatial light spot parameters. Subsequently, calcium fluorescence image sequences are acquired, so that the reserve pool state vector reflects more the intrinsic propagation response of the in vitro neural network to the image input, rather than the local artifacts formed by direct light stimulation. Finally, the trained output layer outputs the image recognition result based on the reserve pool state vector, thereby achieving image classification with a lower training cost for the output layer.
[0008] In some embodiments of the present invention, in the steps of acquiring an image sample to be identified, performing size normalization and grayscale normalization on the image sample to be identified, and obtaining an image grayscale matrix, the size of the image sample to be identified is scaled to M×N pixels, and the pixel value of each pixel is mapped to a grayscale value within a preset grayscale range, thereby obtaining an image grayscale matrix composed of M rows and N columns of grayscale values.
[0009] In the specific implementation process, M and N are both integers greater than or equal to 8 and less than or equal to 64.
[0010] In some embodiments of the present invention, in the step of determining the inner radius and spot width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, the inner radius and spot width of the annular light spot are determined based on the gray value of the pixel using the following formula: in, Represents pixels The inner radius of the corresponding annular light spot; Indicates the preset maximum radius; Represents pixels grayscale value, Represents pixels The width of the corresponding annular light spot.
[0011] In some embodiments of the present invention, in the step of determining the inner radius and spot width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, each annular light spot has a target position set at its center. After determining the inner radius and spot width of the annular light spot, the center of each annular light spot is matched to the corresponding target position to obtain the stimulation pattern.
[0012] In some embodiments of the present invention, in the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to a preset column stimulation period based on the stimulation sequence, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern, the light stimulation of the stimulation patterns corresponding to multiple region vectors is applied to the in vitro neuronal network reserve pool layer according to the stimulation sequence using a preset column stimulation period.
[0013] In some embodiments of the present invention, in the step of applying the light stimulation of the stimulation patterns corresponding to multiple region vectors to the in vitro neuronal network reserve pool layer according to the stimulation sequence using a preset column stimulation cycle, the light stimulation of the stimulation patterns corresponding to multiple region vectors is performed for the effective illumination period according to the stimulation sequence, and after the light stimulation of all region vectors of the image grayscale matrix is completed, the light-free recovery period is entered.
[0014] In the specific implementation process, the next light stimulation of the image grayscale matrix is performed after the light-free recovery period.
[0015] In some embodiments of the present invention, in the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to a preset stimulation cycle based on the stimulation sequence, so that the in vitro neuronal network reserve pool layer generates a calcium fluorescence response to the stimulation pattern, the in vitro neuronal network reserve pool layer is a primary mouse cortical neuron cultured in a PDMS micro-region chamber. After viral infection, the primary mouse cortical neuron co-expresses the light-sensitive protein CheRiff and the red calcium indicator jRCaMP1b, enabling the primary mouse cortical neuronal network to generate an optogenetic response to the light stimulation pattern and output a calcium fluorescence response.
[0016] In some embodiments of the present invention, in the step of acquiring the calcium fluorescence response generated by the in vitro neuronal network reserve pool layer through a calcium imaging reading module equipped with a camera and constructing a calcium fluorescence response sequence, the calcium fluorescence response image of the in vitro neuronal network reserve pool layer is acquired by the camera, and the real-time fluorescence response value of each neuron is determined from the calcium fluorescence response image based on the pre-labeled neuronal target position of the in vitro neuronal network reserve pool layer, and the calcium fluorescence response sequence is obtained by combining the real-time fluorescence response values at multiple time points.
[0017] In some embodiments of the present invention, in the step of determining the relative fluorescence change sequence based on the calcium fluorescence response sequence of each candidate neuron, the relative fluorescence change value is calculated based on the real-time fluorescence response value and baseline fluorescence intensity of each neuron in the calcium fluorescence response sequence, and the relative fluorescence change value of each neuron at multiple time points is constructed as a relative fluorescence change sequence.
[0018] A second aspect of the present invention also provides an image recognition system based on an in vitro neural network reservoir computing architecture. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.
[0019] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the text, or may be learned by practice of the invention. The objects and other advantages of the invention will become apparent from the description and the accompanying drawings.
[0020] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description
[0021] The accompanying drawings, which are provided to further illustrate the invention and form part of this application, are not intended to limit the scope of the invention.
[0022] Figure 1 This is a schematic diagram of the processing flow of one implementation of this solution; Figure 2 This is a schematic diagram of the overall architecture of one implementation of the solution; Figure 3 This is a schematic diagram of the optical processing architecture of this solution; Figure 4 This is a schematic diagram of the in vitro neuronal network reservoir layer and its response; Figure 5 A schematic diagram illustrating the encoding of grayscale value distribution into corresponding optical stimulus patterns; Figure 6 This is a schematic diagram of the light stimulation rules in this scheme; Figure 7 This is a schematic diagram showing the calcium activity response under light stimulation conditions represented by different categories of numbers; Figure 8 This is a schematic diagram of the classification results of MNIST data obtained by stimulation of the in vitro neural network reserve pool layer. Figure 9 This is a diagram illustrating the comparison of classification performance. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.
[0024] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0025] like Figure 1 and 2 As shown, this invention proposes an image recognition method based on an in vitro neural network reservoir computing architecture. The steps of the method include: Step S100: Obtain the image sample to be recognized, and perform size normalization and grayscale normalization on the image sample to be recognized to obtain the image grayscale matrix; In the specific implementation process, after acquiring the image sample to be identified, the image sample is first normalized in size, uniformly scaling it to M×N pixels. M represents the number of rows in the image grayscale matrix, and N represents the number of columns in the image grayscale matrix. Both M and N can be determined based on the field of view of the neuron culture, the spatial resolution of the digital micromirror device, and the complexity of the task.
[0026] In one implementation, M and N can each take integer values between 8 and 64; in a handwritten digit recognition task, M=20 and N=20 can be used as an example.
[0027] After size normalization, grayscale normalization is performed on each pixel to obtain grayscale values within a preset grayscale range. If the original image is a color image, it can be converted to a grayscale image first, and then normalization can be performed; if the original image is already a grayscale image, its grayscale values are directly normalized. The image grayscale matrix is formed by M rows and N columns of normalized grayscale values.
[0028] Specifically, the structural order is either column order from left to right or row order from top to bottom. If the structural order is column order from left to right, then the region vector is a column vector; if the structural order is row order from top to bottom, then the region vector is a row vector.
[0029] If a left-to-right column order is used, multiple column vectors are generated according to the column order of the image grayscale matrix. The nth column vector is formed by arranging the M grayscale values in the nth column of the image grayscale matrix in row order, where n is an integer greater than or equal to 1 and less than or equal to N. These column vectors serve as inputs for subsequent column stimulus patterns, enabling the column-oriented structure of the two-dimensional image to enter the in vitro neural network reserve pool layer in a temporal sequence.
[0030] Step S200: Divide the image grayscale matrix into multiple region vectors, and determine the stimulus order of the region vectors based on the structural order of the image grayscale matrix; For any given region vector, generate a corresponding column stimulus pattern based on each gray value in the region vector.
[0031] In one implementation, the maximum radius can be preset.
[0032] When generating column stimulation patterns, multiple annular light spots are arranged sequentially in the pattern plane of the digital micromirror device according to the position of their corresponding pixels in the column direction. Then, the N column stimulation patterns are arranged in structural order to obtain the photostimulation pattern.
[0033] Step S300: Determine the inner radius and spot width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combine the annular light spots of the region vector to obtain the stimulation pattern of the region vector. Step S400: Based on the stimulation sequence, light stimulation is applied to the preset in vitro neuronal network reserve pool layer according to the preset column stimulation cycle, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern. Specifically, applying light stimulation using a light stimulation input layer can include a light source with a center wavelength of 470 nm, a bandpass filter of 450 nm to 490 nm, a total internal reflection prism, a digital micromirror device (DMM), a dichroic mirror, a lens, and an objective lens. Blue light emitted from the light source is filtered by the bandpass filter and then enters the DMM. The DMM spatially modulates the incident light according to a column stimulation pattern. The modulated beam is then projected through the dichroic mirror, lens, and objective lens onto the stimulation region in the in vitro neuronal network reserve pool layer.
[0034] In one embodiment, the light stimulation intensity can be set from 0.5 mW / mm² to 1.0 mW / mm².
[0035] Step S500: The calcium fluorescence response generated by the in vitro neuronal network reserve pool layer is acquired by a calcium imaging reading module equipped with a camera, and the real-time fluorescence response value is obtained. like Figure 3 As shown, Figure 3 In this module, LED stands for Light Emitting Diode; F for a filter; TIR for a Total Internal Reflection Prism; TL for a Lens; M for a Reflector; DM for a Dichroic Mirror; L for a Lens; OBJ for an Objective Lens; and DMD for a Digital Micromirror Device. The calcium imaging readout module may include a fluorescence excitation source with a center wavelength of 565 nm, a 545 nm to 585 nm filter, an Objective Lens, a long-pass filter, and a camera. Light emitted from the fluorescence excitation source is filtered and then applied to the in vitro neuronal network reserve layer, causing neurons expressing jRCaMP1b to generate calcium fluorescence signals. These signals are collected by the Objective Lens and then pass through a long-pass filter with a starting wavelength of 590 nm before entering the camera. The camera can acquire calcium fluorescence image sequences at a frame rate of 10 frames per second or other frame rates that meet the requirements for recording neuronal activity.
[0036] After obtaining the calcium fluorescence image sequence, pre-labeled neuronal target locations are identified from the signal reading region, and real-time fluorescence response values are extracted for each neuronal target location.
[0037] In the specific implementation process, the in vitro neuronal network reserve pool layer can be formed by culturing primary mouse cortical neurons in polydimethylsiloxane (PDMS) microcompartments. The PDMS microcompartments are used to define the neuronal culture area, ensuring that the spatial scale of the neuronal network matches the field of view for subsequent light stimulation input and calcium imaging readout. The size of the microcompartment can be set according to the image recognition task and the field of view for microscopic imaging, for example, 1.5 mm × 1.0 mm. After 14 days of in vitro culture, the neuronal network reaches a functionally mature state. Motion correction algorithms are used to compensate for spatial displacement of the image sequence during acquisition, and spatial filtering techniques are combined to filter out background noise to improve the image signal-to-noise ratio. Adaptive thresholding and machine learning algorithms are used to automatically identify neuronal activity characteristics, and neurons are initially segmented based on morphological structure. Specifically, segmentation can be done manually or through machine learning to identify and segment the neurons, thus obtaining the target location region of the neuron.
[0038] After the neuronal network reached a culture stage where it could stably generate spontaneous calcium activity, a viral infection method was used to induce co-expression of the light-sensitive protein CheRiff and the red calcium indicator jRCaMP1b in the neurons. CheRiff was used to enable the neurons to produce an optogenetic response to blue light stimulation, and jRCaMP1b was used to enable the neuronal activity to be read in the form of red calcium fluorescence. This is because the stimulation spectrum of the light-sensitive protein and the imaging spectrum of the red calcium indicator...
[0039] The in vitro neural network reserve pool layer not only serves as a passive image response object but also participates in image information mapping as a physical reserve pool. After being encoded by the input layer, the image sample to be identified enters the in vitro neural network reserve pool layer through light stimulation. The synaptic connections between neurons and network dynamics map the input information into a high-dimensional calcium fluorescence response state. The subsequent output layer can complete the category readout without training the internal connection weights of the in vitro neural network.
[0040] Step S600: Determine the relative fluorescence change sequence based on the calcium fluorescence response sequence of each candidate neuron, input the relative fluorescence change sequence into a preset classification model, and output the classification result through the classification model.
[0041] Using the above scheme, the image sample to be identified is first converted into an image grayscale matrix. Then, the image grayscale matrix is converted into a spatiotemporal sequence of light stimulation patterns in the form of ring-shaped light spots according to the structural order. This allows the spatial grayscale changes of the two-dimensional image to enter the in vitro neural network reserve pool layer in both temporal order and spatial light spot parameters. Subsequently, calcium fluorescence image sequences are acquired, so that the reserve pool state vector reflects more the intrinsic propagation response of the in vitro neural network to the image input, rather than the local artifacts formed by direct light stimulation. Finally, the trained output layer outputs the image recognition result based on the reserve pool state vector, thereby achieving image classification with a lower training cost for the output layer.
[0042] In some embodiments of the present invention, in the steps of acquiring an image sample to be identified, performing size normalization and grayscale normalization on the image sample to be identified, and obtaining an image grayscale matrix, the size of the image sample to be identified is scaled to M×N pixels, and the pixel value of each pixel is mapped to a grayscale value within a preset grayscale range, thereby obtaining an image grayscale matrix composed of M rows and N columns of grayscale values.
[0043] In the specific implementation process (Preferred value: M=N=20).
[0044] like Figure 5 As shown, in some embodiments of the present invention, in the step of determining the inner radius and spot width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, the inner radius and spot width of the annular light spot are determined based on the gray value of the pixel using the following formula: in, Represents pixels The inner radius of the corresponding annular light spot; Indicates the preset maximum radius; Represents pixels grayscale value, Represents pixels The width of the corresponding annular light spot.
[0045] Specifically, The value is 28~35 μm, preferably 32 μm; 255 is the maximum grayscale value.
[0046] Using the above scheme, the pixel grayscale value is mapped to the inner radius and width of the annular light spot. The annular light spot is then used to perform optogenetic stimulation on the in vitro neural network. Since the annular light spot can distribute light energy within the annular region rather than concentrating it at the center, it can reduce the phototoxicity and response saturation risk caused by excessive local illumination while inducing an effective response in the neural network. At the same time, the inner radius and width of the annular light spot can vary with the pixel grayscale value, so that the grayscale difference of the two-dimensional image is transformed into a light stimulation pattern with spatial scale differences. This enhances the dynamic response differences formed in the biological neural network by different image inputs, and improves the stability and distinguishability of subsequent reservoir state extraction and image recognition classification.
[0047] In some embodiments of the present invention, in the step of determining the inner radius and width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, each annular light spot has a target position set at its center, such as... Figure 4 As shown in (a), after determining the inner radius and width of the annular light spot, the center of each annular light spot is matched to the corresponding target position to obtain a stimulation pattern.
[0048] In the specific implementation process, the annular region of the annular light spot is subjected to light stimulation.
[0049] In some embodiments of the present invention, in the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to a preset column stimulation period based on the stimulation sequence, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern, the light stimulation of the stimulation patterns corresponding to multiple region vectors is applied to the in vitro neuronal network reserve pool layer according to the stimulation sequence using a preset column stimulation period.
[0050] In some embodiments of the invention, the stimulus order is the horizontal left-to-right order of the column vectors.
[0051] like Figure 6 As shown, in some embodiments of the present invention, in the step of applying the light stimulation of the stimulation patterns corresponding to multiple region vectors to the in vitro neuronal network reserve pool layer according to the stimulation order using a preset column stimulation cycle, the light stimulation of the stimulation patterns corresponding to multiple region vectors is performed for the effective illumination period according to the stimulation order, and after completing the light stimulation of all region vectors of the image grayscale matrix, the light-free recovery period is entered.
[0052] Specifically, the total duration of each time-series signal is 500 ms, of which the effective light stimulation duration is controlled within 100–200 ms, and a 5-second light-free recovery period is set between adjacent samples.
[0053] In the specific implementation process, the next light stimulation of the image grayscale matrix is performed after the light-free recovery period.
[0054] In some embodiments of the present invention, in the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to a preset stimulation cycle based on the stimulation sequence, so that the in vitro neuronal network reserve pool layer generates a calcium fluorescence response to the stimulation pattern, the in vitro neuronal network reserve pool layer is a primary mouse cortical neuron cultured in a PDMS micro-region chamber. After viral infection, the primary mouse cortical neuron co-expresses the light-sensitive protein CheRiff and the red calcium indicator jRCaMP1b, enabling the primary mouse cortical neuronal network to generate an optogenetic response to the light stimulation pattern and output a calcium fluorescence response.
[0055] In some embodiments of the present invention, in the step of acquiring the calcium fluorescence response generated by the in vitro neuronal network reserve pool layer through a calcium imaging reading module equipped with a camera and constructing a calcium fluorescence response sequence, the calcium fluorescence response image of the in vitro neuronal network reserve pool layer is acquired by the camera, and the real-time fluorescence response value of each neuron is determined from the calcium fluorescence response image based on the pre-labeled neuronal target position of the in vitro neuronal network reserve pool layer, and the calcium fluorescence response sequence is obtained by combining the real-time fluorescence response values at multiple time points.
[0056] In some embodiments of the present invention, in the step of determining the relative fluorescence change sequence based on the calcium fluorescence response sequence of each candidate neuron, a relative fluorescence change value is calculated based on the real-time fluorescence response value and baseline fluorescence intensity of each neuron in the calcium fluorescence response sequence. A relative fluorescence change sequence is constructed for each neuron at multiple time points, such as... Figure 4 As shown in (b).
[0057] like Figure 7 As shown, in the specific implementation process, the sum of all real-time fluorescence response values within the target location region of each neuron is calculated. F total Because the original signal contains neural signals originating from extrafocal dendrite and axonal activity. F neu The target location of the neuron is a circular region. F neuThe fluorescence response value is the average value of the fluorescence response values in the outward-extending rings of the target area. These fluorescence response values can be brightness values. The width of the rings is the same as the radius of the circular region at the target location. This paper uses a subtraction method to correct for this value, thus obtaining the neuronal cell body fluorescence signal. = F total - k × F neu .
[0058] Based on the optimization of experimental conditions, the correction coefficient was adjusted. k The value was set to 0.8. Finally, the relative fluorescence change was calculated: in, This represents the relative fluorescence change value. Indicates baseline fluorescence intensity. To record the average of the 50% of the lowest fluorescence intensities within the target location region during the first 10 seconds after the start, Indicates the fluorescent signal of the neuronal cell body. This represents the difference between the real-time fluorescence response value and the baseline fluorescence intensity.
[0059] In some embodiments of the present invention, the classification model may employ ridge regression, least squares classifier, linear kernel classifier of support vector machine, or other classification models suitable for the reservoir calculation output layer.
[0060] In one implementation, the output layer weights can be determined using ridge regression, and the calculation can be expressed as follows: in, This represents the output layer weight matrix; X This represents the state matrix composed of the state vectors of the training reservoir; Y This represents a label matrix consisting of the category labels of the training image samples; λ Represents the regularization coefficient; I This represents the identity matrix. The result obtained through the above calculations... This is used to map the reservoir state vector to the category scores corresponding to different image categories.
[0061] During the recognition phase, the classification model can obtain multiple category scores. The biometric computing system determines the image category with the highest category score as the image recognition result. For handwritten digit recognition tasks, the image category can be any digit category from 0 to 9; for other image recognition tasks, the image category can be set according to the labeling system of the training data. Figure 8 and 9This diagram illustrates the experimental results for comparing the existence of an in vitro neural network reserve pool based on this scheme.
[0062] This invention uses biological neural networks as a physical reservoir, fully utilizing their inherent nonlinear dynamic characteristics to automatically extract and map high-dimensional image features. Through systematic investigation of the effects of light stimulation intensity and pulse duration on neuronal responses, the optimal combination of optical parameters is determined. When the light stimulation intensity is 0.75 mW / mm² and the pulse duration is 200 ms, the neural network exhibits the most significant differences in response patterns to different numbers. Figure 8 As shown, the state space of the reservoir is sufficiently enriched, and the class separability is significantly improved. The collected calcium fluorescence signals are mapped to reservoir states and then input into a ridge regression classifier for training. The output weight matrix is optimized, and finally, a "winner-takes-all" strategy is used to predict the digit class. This invention does not limit the ratio of training and test sets; it can be flexibly set according to actual task requirements. In the MNIST handwritten digit 10 classification task, as an example, 80 digits of each class are selected as the training set and 20 as the test set. The system constructed in this invention achieves a recognition accuracy of 75.3%, which is higher than the accuracy of purely linear classification of MNIST (70.6%). Figure 8 As shown, both the Macro-F1 value and the Cohen-Kappa coefficient exhibit a consistent upward trend, such as... Figure 9 As shown, this fully verifies the significant advantages of biological reservoirs in nonlinear mapping and high-dimensional feature extraction.
[0063] This invention also provides an image recognition system based on an in vitro neural network reservoir computing architecture. The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method described above.
[0064] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.
[0065] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.
[0066] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.
[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An image recognition method based on an in-vitro neuronal network reservoir computing architecture, characterized in that, The steps of the method include: Obtain the image sample to be identified, and perform size normalization and grayscale normalization on the image sample to be identified to obtain the image grayscale matrix; The image grayscale matrix is divided into multiple region vectors, and the stimulation order of the region vectors is determined based on the structural order of the image grayscale matrix. The inner radius and width of the corresponding annular light spot are determined based on the gray value of each pixel in the region vector. The annular light spots of the region vector are combined to obtain the stimulation pattern of the region vector. Based on the stimulation sequence, light stimulation is applied to a preset in vitro neuronal network reserve pool layer according to a preset column stimulation cycle, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern. The calcium fluorescence response generated by the in vitro neuronal network reserve pool layer is acquired by a calcium imaging readout module equipped with a camera, and constructed into a calcium fluorescence response sequence. Based on the calcium fluorescence response sequence of each candidate neuron, a relative fluorescence change sequence is determined, and the relative fluorescence change sequence is input into a preset classification model. The classification model then outputs the classification result.
2. The image recognition method based on an external neural network reservoir computing architecture according to claim 1, characterized in that, In the steps of acquiring the image sample to be identified, performing size normalization and grayscale normalization on the image sample to be identified, and obtaining the image grayscale matrix, the size of the image sample to be identified is scaled to M×N pixels, and the pixel value of each pixel is mapped to a grayscale value within a preset grayscale range, resulting in an image grayscale matrix composed of M rows and N columns of grayscale values.
3. The image recognition method based on an in vitro neural network reservoir computing architecture according to claim 1, characterized in that, In the step of determining the inner radius and width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, the inner radius and width of the annular light spot are determined based on the gray value of the pixel using the following formula: in, Represents pixels The inner radius of the corresponding annular light spot; Indicates the preset maximum radius; Represents pixels grayscale value, Represents pixels The width of the corresponding annular light spot.
4. The image recognition method based on an in vitro neural network reservoir computing architecture according to claim 1, characterized in that, In the step of determining the inner radius and width of the corresponding annular light spot based on the gray value of each pixel in the region vector, and combining the annular light spots of the region vector to obtain the stimulation pattern of the region vector, each annular light spot has a target position set at its center. After determining the inner radius and width of the annular light spot, the center of each annular light spot is matched to the corresponding target position to obtain the stimulation pattern.
5. The image recognition method based on an external neural network reservoir computing architecture according to claim 1, characterized in that, In the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to the stimulation sequence and a preset column stimulation cycle, so that the in vitro neuronal network reserve pool layer generates calcium fluorescence response to the stimulation pattern, the light stimulation of the stimulation patterns corresponding to multiple region vectors is applied to the in vitro neuronal network reserve pool layer according to the stimulation sequence and a preset column stimulation cycle.
6. The image recognition method based on an in vitro neural network reservoir computing architecture according to claim 5, characterized in that, In the step of applying the light stimulation of the stimulation patterns corresponding to multiple region vectors to the in vitro neuronal network reserve pool layer according to the stimulation order using a preset column stimulation cycle, the light stimulation of the stimulation patterns corresponding to multiple region vectors is performed for the effective illumination period according to the stimulation order, and after the light stimulation of all region vectors of the image grayscale matrix is completed, the light-free recovery period is entered.
7. The image recognition method based on an in vitro neural network reservoir computing architecture according to claim 1, characterized in that, In the step of applying light stimulation to a preset in vitro neuronal network reserve pool layer according to a preset stimulation cycle based on the stimulation sequence, so that the in vitro neuronal network reserve pool layer generates a calcium fluorescence response to the stimulation pattern, the in vitro neuronal network reserve pool layer is a primary mouse cortical neuron cultured in a PDMS micro-region chamber. After viral infection, the primary mouse cortical neuron co-expresses the light-sensitive protein CheRiff and the red calcium indicator jRCaMP1b, enabling the primary mouse cortical neuronal network to generate an optogenetic response to the light stimulation pattern and output a calcium fluorescence response.
8. The image recognition method based on an external neural network reservoir computing architecture according to claim 1, characterized in that, In the step of acquiring the calcium fluorescence response generated by the in vitro neuronal network reserve pool layer through a calcium imaging reading module equipped with a camera and constructing a calcium fluorescence response sequence, the calcium fluorescence response image of the in vitro neuronal network reserve pool layer is acquired by the camera, and the real-time fluorescence response value of each neuron is determined from the calcium fluorescence response image based on the pre-labeled neuronal target position of the in vitro neuronal network reserve pool layer. The calcium fluorescence response sequence is obtained by combining the real-time fluorescence response values at multiple time points.
9. The image recognition method based on an in vitro neural network reservoir computing architecture according to claim 1, characterized in that, In the step of determining the relative fluorescence change sequence based on the calcium fluorescence response sequence of each candidate neuron, the relative fluorescence change value is calculated based on the real-time fluorescence response value and baseline fluorescence intensity of each neuron in the calcium fluorescence response sequence, and the relative fluorescence change value of each neuron at multiple time points is constructed as a relative fluorescence change sequence.
10. An image recognition system based on an external neural network reservoir computing architecture, characterized in that, The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps of the method as described in any one of claims 1 to 9.