Optical high-parallel matrix addition framework based on multi-imaging projection and calculation method of optical high-parallel matrix addition framework
By using an optical high-parallel matrix addition architecture based on multi-imaging projection and a photoelectric hybrid system, low-energy-consumption, high-speed, high-computing-power matrix addition operations are realized, solving the problem of low computing efficiency of traditional CPUs and breaking through the bottleneck of electronic computing.
Patent Information
- Application Number
- CN202511443802.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-02-06
AI Technical Summary
In existing technologies, matrix operations are inefficient on traditional serial CPUs, resulting in a sharp increase in computation time and energy consumption, making it difficult to meet the needs of large-scale data computation.
An optical high-parallel matrix addition architecture based on multi-imaging projection is adopted. Using a hybrid optoelectronic system, parallel addition operations of matrix elements are realized through optical signal beam splitting and diffraction imaging technology. It includes a light source signal loading module, an imaging projection module, an optical signal detection module, and an addition calculation acceleration module. Dammann gratings and Fourier lenses are used for beam splitting and imaging projection of matrix information.
It achieves low-energy, high-speed, high-computing-power matrix addition operations, and the calculation results can be transferred between optical platforms. The calculation efficiency is improved by a thousand times, the power consumption is significantly reduced, and the calculation accuracy meets the requirements of AI models.
Smart Images
Figure CN121478080A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of optical computing, and particularly relates to an optical high-parallel matrix addition framework based on multi-imaging projection and a calculation method thereof. BACKGROUND
[0002] With the rapid development of artificial intelligence technology represented by deep learning, it marks that human beings have entered the big data era. The growth rate of the increasing data calculation demand of the society far exceeds the Moore's Law of the supply of computing power. Therefore, in order to meet the needs of social production, it is urgent to find a breakthrough point of computing power. Countries around the world are working to build a computing system framework with higher computing power and lower power consumption. The core three elements of artificial intelligence include computing power, data and algorithm. Looking at the computing power alone, whether in the field of deep learning or in the field of image processing, a large number of matrix operations are involved. Matrix operations account for more than 80% of the calculation amount. Matrix operations are extremely important, but the operation is very low in efficiency on the traditional serial CPU. With the characteristics of photonic high parallelism, low energy consumption and fast action speed, replacing electronic computing with a dedicated hardware accelerator has again entered people's field of vision. In theory, the light signal carrying data runs at the speed of light in transmission. Such characteristics make it possible to design a super-low delay computing platform with optical structure. Secondly, the modulation mode of light wave, such as wavelength, phase, polarization and amplitude, provides rich multiplexing degrees of freedom for optical domain computing while greatly improving the growth rate of computing power. In addition, compared with electrons, photons have no ohmic loss and thus have lower energy consumption. Moreover, the higher bandwidth provided by photons is also very beneficial to the acceleration of neural networks. Of course, the current all-optical computing technology and phase correlator are not mature enough. The optical-electric hybrid intelligent computing system framework is the mainstream trend of the current optical computing system. By developing appropriate coding, parallel algorithms and architecture system, the parallel characteristics of optics are fully utilized. Photons are expected to achieve higher computing power density and energy efficiency ratio than electrons in the same unit area, and higher computing accuracy in the photonic system.
[0003] As a hardware accelerator for addition, it can be used to process massive linear data. According to the parallel characteristics of photons, its computing power grows in N 2In processing large-scale linear data analysis based on statistics, image filtering and transformation and enhancement, and signal linear transformation processing has obvious advantages. At the same time, the electronic computer executes the serial execution instruction, and the obtained result needs to be stored in sequence, and the next operation is executed in time sequence. When the calculation scale becomes larger, the processing time and energy consumption will increase sharply, which is the key factor restricting the performance of electronic computer. The optical method has ultra-low operation time consumption, serial processing of each operation, only time delay of data upload and download, no intermediate calculation result, and no need to occupy additional storage resources and processing time. The calculation result can be transferred to other platforms at will. The overall calculation efficiency of the optoelectronic hybrid intelligent computing system will be improved by thousands of times, the power consumption will be greatly reduced, the calculation result can have good tolerance, and the calculation accuracy can meet part of the demand of AI model. Of course, when the matrix size N is large enough, the calculation power is expected to be comparable to that of electronic computer, and the energy consumption and high parallel data processing characteristics are unmatched by electronic computer. SUMMARY
[0004] The main purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide an optical matrix addition calculation accelerator based on multi-imaging projection and a calculation method thereof, which aims to fully utilize the parallel characteristics of photons through an optoelectronic hybrid framework to realize low-energy-consumption, high-speed and large-calculating-power matrix addition operation.
[0005] In order to achieve the above purpose, the present application adopts the following technical scheme:
[0006] In a first aspect, the present application provides an optical high-parallel matrix addition framework based on multi-imaging projection, comprising a light source, a light source signal loading module, an imaging projection module, an optical signal detection module and an addition calculation acceleration module.
[0007] The light source signal loading module is used for loading the to-be-modulated optical signal, and realizing the function of selecting a specific matrix for different scale operation.
[0008] The imaging projection module is used for diffracting the matrix information carrying the optical signal into different angle sub-beams through a Dammann grating beam splitter, and imaging and projecting the different matrix elements accumulated between the diffraction orders onto the detection camera at the rear end through a Fourier lens, so as to project the addition result of the matrix elements onto the detection plane.
[0009] The optical signal detection module is used for recording the linear operation result of the matrix elements according to the size of the gray scale, and deducting the background noise generated by the related device, so as to realize the detection of the addition result of the to-be-calculated optical matrix gray scale.
[0010] The addition calculation acceleration module is used for matching the size of the optical matrix element with the sliding relationship between the adjacent two diffraction orders, so as to realize the addition of different size matrices and optical parallel linear operation.
[0011] As a preferred technical solution, the light source is an LED light source, a vertical cavity surface-mount semiconductor laser array, a digital projection processor, or a fiber array.
[0012] As a preferred technical solution, the light source signal loading module includes a first plano-convex lens, a 4f system, an aperture, a polarization beam splitter, and a spatial light modulator. The light emitted by the light source is diffused into quasi-parallel light by the first plano-convex lens. The quasi-parallel light is further formed into a standard collimated beam by the 4f system. The collimated beam is reflected by the polarization beam splitter to the spatial light modulator. The spatial light modulator loads grayscale information and converts the optical information into numerical information through encoding. The modulated collimated beam is reflected by the liquid crystal and its polarization state is changed before it is transmitted through the polarization beam splitter to complete the signal loading.
[0013] As a preferred technical solution, the 4f system includes a second plano-convex lens and a third plano-convex lens arranged opposite to each other, with an aperture stop disposed between the second plano-convex lens and the third plano-convex lens.
[0014] As a preferred technical solution, the imaging projection module includes a first Fourier lens, a Dammann grating, and a second Fourier lens. The light signal emitted from the polarization beam splitter is focused into a light spot by the first Fourier lens, which facilitates the diffraction of the Dammann grating. The Dammann grating splits the matrix carrying the information to be calculated into multiple sub-beams with the same light intensity information at different diffraction angles. The sub-beams are imaged and projected onto the detection plane by the second Fourier lens at a specific diffraction angle.
[0015] As a preferred technical solution, the addition calculation acceleration module adjusts the distance between the Damman grating and the first Fourier lens and the second Fourier lens, and the size of the matrix pixels loaded by the light source signal loading module, so that the light modulated by the Damman grating is split into beams according to a set diffraction angle, and adjacent diffraction orders slide a set distance from each other to achieve the superposition of weight values of different matrix elements.
[0016] The diffraction angle is affected by the diffraction formula, as follows:
[0017] ;
[0018] The diffraction order sliding step size is affected by the following formula:
[0019] ;
[0020] in, , These refer to the Daman gratings in sequence. The diffraction angle and grating period of each diffraction order. It is the distance from the Dammann grating to the second Fourier lens. ;
[0021] The sliding step size s is matched with the matrix cell size ∆, and matrix addition operations of different sizes are achieved by adjusting the ratio between s and ∆.
[0022] As a preferred technical solution, the Damman grating is an even-numbered Damman grating, for a size of The optical matrix is split by a Dammann beam, where one of the diffraction orders is (+1, +1). The matrix elements from row 11 to row 20 are superimposed on the elements in row 1 and row 10 of the matrix with diffraction order (+1, -1). In the transverse diffraction direction (+1, -1), the elements in row 11 are superimposed on the elements in row 10 of the matrix with diffraction order (+1, -1). ) and (-1, The diffraction order is exactly offset by 10 elements, thus achieving... Matrix addition yields a matrix .
[0023] As a preferred technical solution, the optical signal detection module includes a high dynamic range and high sensitivity camera and a grayscale encoding unit; the camera is a CCD, CMOS or sCMOS camera, used to receive the calculation results of a specific planar grayscale matrix; the grayscale encoding unit realizes natural number addition operations through grayscale quantization and assignment, and can adapt to the needs of positive and negative number and complex number addition calculations.
[0024] Secondly, the present invention provides a computational method based on a multi-imaging projection optical highly parallel matrix addition architecture, comprising the following steps:
[0025] Light source signal loading: The light source signal loading module loads the matrix information to be calculated onto the collimated beam in grayscale encoding form to complete the optical signal modulation;
[0026] Multiple imaging projection: The beam carrying matrix information is split into multiple sub-beams containing the same information by using the Dammann grating in the imaging projection module. The sub-beams are then imaged and projected onto the detection plane by Fourier lenses.
[0027] Additive acceleration control: The spacing between the Dammann grating and the Fourier lens and the size of the matrix pixels are adjusted by the additive calculation acceleration module, so that adjacent diffraction order sub-beams slide to a preset distance, realizing the spatial superposition of matrix elements;
[0028] Signal detection and decoding: The optical signal detection module receives the superimposed grayscale matrix signal, subtracts background noise, and obtains the digital result of matrix addition through grayscale decoding.
[0029] As a preferred technical solution, the additive acceleration control step further includes the following steps:
[0030] The sliding distance s of adjacent diffraction order sub-beams and the matrix element size Delta satisfy s=k*Delta, k is a positive integer, the preset k value is used to realize the weighted superposition of matrix elements, when k=1, the direct accumulation of adjacent matrix elements is realized, when k=n, the accumulation of matrix elements with interval n-1 elements is realized, and the sliding distance s is adjusted through the interval d of the Dammann grating and the Fourier lens and the diffraction angle theta, and s=d*tan(theta) is satisfied.
[0031] Compared with the prior art, the present application has the following advantages and beneficial effects:
[0032] The present application loads multiple sets of linear data and arranges them in the form of a matrix, and uses a diffraction beam splitting device to realize linear operation of the matrix elements in an optical manner. The optoelectronic hybrid computing architecture has good parallel processing capability, and the computing results can be well transferred to other platforms. Secondly, since optical computing has almost no delay, the time consumption of the optoelectronic hybrid intelligent computing system is mainly in the uploading and downloading of data, so the computing time consumption is much lower than that of an electronic computer. Essentially, it uses spatial parallel operation to exchange time acceleration. In addition, the computing power growth speed of the present application is , and with the growth of the matrix scale , the calculation increases sharply, but the energy consumption does not increase significantly. The present application has obvious advantages in processing massive linear data and matrix operation. BRIEF DESCRIPTION OF DRAWINGS
[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0034] Figure 1 The present application is an optical high-parallel matrix addition architecture based on multiple imaging projection.
[0035] Figure 2 The present application is an addition calculation acceleration module.
[0036] Figure 3 The present application is a multiple imaging projection module.
[0037] Figure 4 The present application is an optical signal detection module.
[0038] Explanation of the reference numerals:
[0039] 101-Light source; 201-Light source signal loading module; 2011-First plano-convex lens; 2012-Second plano-convex lens; 2013-Aperture stop; 2014-Third plano-convex lens; 2015-Spatial light modulator; 2016-Polarization beam splitter; 301-Imaging projection module; 3011-First Fourier lens; 3012-Damman grating; 3013-Second Fourier lens; 401-Optical signal detection module. Detailed Implementation
[0040] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.
[0041] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.
[0042] like Figure 1 As shown, this embodiment provides an optical high-parallel matrix addition architecture based on multi-imaging projection, including a light source 101, a light source signal loading module 201, an imaging projection module 301, an optical signal detection module 401, and an addition calculation acceleration module. First, the light source 101 emits an initial light signal, which enters the light source signal loading module 201. In this module, the light signal is processed by optical elements to form a collimated beam. Then, a spatial light modulator loads the grayscale information of the matrix to be calculated, completing the modulation of the optical signal so that it carries matrix data information. Next, the light signal carrying the matrix information enters the imaging projection module 301. A Damman grating diffracts this beam into multiple sub-beams containing the same information. Each sub-beam propagates at a different diffraction angle and is then imaged and projected onto the detection plane at a specific angle by a Fourier lens, realizing the spatial projection of the matrix element accumulation result. Subsequently, the optical signal detection module 401 receives the light signal projected onto the detection plane, records the grayscale result of the linear operation of matrix elements through a high-sensitivity camera, and the grayscale encoding unit subtracts background noise, performs grayscale quantization and assignment, and converts the optical signal into a digital addition result.
[0043] Further, the light source 101 is selected as an LED light source; of course, to adapt to the needs of different application scenarios, the light source can be replaced by a vertical cavity surface emitting laser array (VCSEL), a digital projection processor (DLP), or a fiber array, which can more directly generate a structured light beam and promote system integration.
[0044] Further, the light source signal loading module 201 is used to load the optical signal to be modulated to realize the operation function of different scales by selecting a specific matrix; it includes a first plano-convex lens 2011, a 4f system, a light barrier 2013, a polarization beam splitter 2015, and a spatial light modulator 2016; the light emitted by the light source is diffused into parallel light by the first plano-convex lens 2011, the parallel light is further formed into a standard collimated light beam by the 4f system, the collimated light beam is reflected to the spatial light modulator 2016 by the polarization beam splitter 2015, the spatial light modulator loads the gray scale information and converts the optical information into numerical information through coding, and the collimated light beam after being modulated is reflected by the liquid crystal and changes the polarization state, and then transmits through the polarization beam splitter 2015 to complete signal loading.
[0045] Further, the 4f system includes a second plano-convex lens 2012 and a third plano-convex lens 2014, and the light barrier is arranged between the second plano-convex lens 2012 and the third plano-convex lens 2014. The second plano-convex lens 2012 receives the light beam emitted from the first plano-convex lens and converts it into a parallel light beam; the third plano-convex lens 2014 performs secondary calibration on the parallel light beam to ensure that the light beam remains parallel during propagation, reduces energy loss and signal distortion caused by light beam divergence, and provides a stable collimated light field for accurate modulation of the subsequent spatial light modulator. The light barrier 2013 located between the two lenses (i.e. the Fourier plane of the 4f system) can filter high-frequency noise and stray light in the light beam. By selecting a light barrier with a suitable aperture, interference signals introduced during light beam propagation (such as incoherent components of the light source, scattered light of optical elements, etc.) can be effectively removed, and the base frequency component carrying effective information can be retained, thereby improving the signal-to-noise ratio and purity of the optical signal.
[0046] Further, the imaging projection module 301 is used to record the linear operation results of the matrix elements according to the size of the gray scale and deduct the background noise generated by the related devices to realize the addition result of the detected optical matrix gray scale to be calculated; it includes a first Fourier lens 3011, a Dammam grating 3012, and a second Fourier lens 3013, and the light signal emitted by the polarization beam splitter passes through the first Fourier lens and converges into a light spot at the focal point, thereby facilitating diffraction by the Dammam grating 3012. The Dammam grating 3012 divides the matrix carrying the to-be-calculated information into multiple sub-beams containing the same light intensity information at different diffraction angles, and the sub-beams are imaged and projected on the detection plane according to a specific diffraction angle by the second Fourier lens.
[0047] The addition calculation acceleration module is used to match the size of the optical matrix elements with the sliding relationship between adjacent diffraction orders, enabling addition of matrices of different sizes and optical parallel linear operations. By adjusting the spacing between the Dammann grating and the Fourier lens, and the size of the matrix pixels, the module controls the sliding distance of adjacent diffraction order sub-beams to match the requirements for matrix element superposition, achieving parallel addition operations for matrices of different sizes. The entire process is completed through the propagation and modulation of optical signals, achieving high parallelism in matrix addition.
[0048] like Figure 2 The diagram shown illustrates the principle of Damman grating multi-imaging and matrix addition in this embodiment. It includes a matrix containing numerical information and a size of... The matrix is divided into 4 sub-beams by the Dammann grating diffraction. Each sub-diffraction order contains all the information of the original matrix. The diffraction angle of the diffraction order is affected by the diffraction formula (1):
[0049] (1)
[0050] Pixel size With adjacent diffraction order sliding step To perform a match, here The diffraction order sliding step size is affected by formula (2):
[0051] (2)
[0052] here , These refer to the Daman gratings in sequence. The diffraction angle and grating period of each diffraction order. It is the distance from the Dammann grating to the Fourier lens. .
[0053] Please refer to it again. Figure 2 In this embodiment, an even-numbered Dammann grating is selected, with a size of... The optical matrix, after being split by the Dammann beam, diffracts into four 20×10 matrices in a rectangular distribution. The middle region between the upper and lower matrices forms a 10×10 overlapping region. The grayscale value of this region corresponds to the matrix addition result, and the boundary of the overlapping region is marked with a dashed box. One of these matrices has a diffraction order of (+1, +1). The matrix elements from row 11 to row 20 are superimposed on the elements in row 1 and row 10 of the matrix with diffraction order (+1, -1). In the transverse diffraction direction, such as (+1, -1)... ) and (-1, The diffraction order is exactly offset by 10 elements, thus achieving... Matrix addition yields a matrix For matrix Each element in ,Right now It's worth noting that this is just a simplified illustration; in principle, more complex implementations are possible. Matrix addition. Of course, diffraction beam-splitting devices with smaller diffraction angles, higher diffraction efficiency, and more uniform beam-splitting ratios can be used, along with Fourier lenses with smaller apertures and focal lengths, and micro-photodetectors with higher dynamic range and more sensitive detection. This fully utilizes the modulator's larger pixel count to achieve larger-scale applications such as... even The matrix addition operation of the size also facilitates the later integration of the system.
[0054] like Figure 3 As shown, the beam carrying weight values a and b output from the spatial light modulator is converged into a spot by Len1 and then incident on the DG. The spot is diffracted and split by the DG, as shown... Figure 3 The 4f system focal plane. The split sub-beams propagate to Len2 and are imaged; adjacent sub-beams intersect at specific positions, such as... Figure 3 The dotted lines intersect. Linear operation can be detected in the detection plane. During imaging projection, the imaging projection module 301 places the Damman grating 3012 at the front focal length of the second Fourier lens 3013, causing adjacent diffraction orders of the grating to slide a certain distance to match the matrix element size. The matrix element pixel size can be adjusted according to the addition scale to achieve addition operations of different sizes, and the calculation result is projected onto the detection plane.
[0055] like Figure 4 As shown, the optical signal detection module 401 is further used to record the linear operation results of the matrix elements according to the gray level and subtract the background noise generated by related devices, so as to detect the addition result of the gray level of the optical matrix to be calculated.
[0056] Furthermore, the optical signal detection module 401 can be a CMOS camera located at the back focal plane of the second Fourier lens 3013, or a CCD camera and an optical signal receiving array, used to detect the linear calculation results of multiple imaging. In addition, a narrowband filter matching the light source to be loaded can be selected to further improve the calculation accuracy. In terms of information processing, 256 gray levels can be quantized and divided into n equal parts, and then all gray values in each region can be assigned a suitable natural number. Addition operations are performed using gray information, and finally, the addition result of natural numbers within a certain range is obtained by decoding the original gray matrix addition result image.
[0057] In another embodiment of the present invention, a method for implementing an optical high-parallel matrix addition architecture based on multiple imaging projection is provided, comprising the following steps:
[0058] S1, light source signal loading: loading the matrix information to be calculated in the form of gray code onto the collimated light beam through a light source signal loading module to complete optical signal modulation;
[0059] S2, multi-imaging projection: using a Dammam grating in the imaging projection module to split the light beam carrying matrix information into multiple sub-beams containing the same information, and the sub-beams are imaged and projected to a detection plane by a Fourier lens;
[0060] S3, addition acceleration regulation: adjusting the distance between the Dammam grating and the Fourier lens and the size of the matrix element through an addition calculation acceleration module, so that the adjacent diffraction order sub-beams slide to a preset distance, realizing the spatial superposition of matrix elements;
[0061] S4, signal detection decoding: the optical signal detection module receives the superimposed gray matrix signal, and obtains the digital result of matrix addition through gray decoding after deducting background noise.
[0062] Further, the addition acceleration regulation step further comprises the following contents:
[0063] The sliding distance s of the adjacent diffraction order sub-beams and the matrix element size Δ satisfy s=k×Δ, k is a positive integer, and the weighted superposition of matrix elements is realized by presetting the value of k, when k=1, the direct accumulation of adjacent matrix elements is realized, and when k=n, the accumulation of matrix elements with an interval of n-1 elements is realized; the sliding distance s is regulated by the distance d between the Dammam grating and the Fourier lens and the diffraction angle θ, and satisfies s=d×tanθ.
[0064] Through the modular design of light source signal loading, multi-imaging projection, optical signal detection and addition calculation acceleration, the application constructs an optical matrix addition calculation accelerator based on multi-imaging projection, fully releases the high parallelism of photons, realizes the matrix addition operation with N² growth of computing power, low energy consumption and high speed. The accelerator effectively breaks through the computing power bottleneck of traditional electronic calculation, provides a new solution for massive matrix operation in the fields of artificial intelligence, neural network and image processing, and has important theoretical value and engineering application prospect.
[0065] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer readable storage medium, and when the program is executed, the processes of the above-mentioned embodiment methods can be included. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0066] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.
[0067] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations and simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods, and are included in the protection scope of the present application.
Claims
1. An optical high-parallel matrix addition architecture based on multiple imaging projection, characterized in that, The light source, a light source signal loading module, an imaging projection module, an optical signal detection module and an addition calculation acceleration module are included. The light source signal loading module is used for loading a to-be-modulated optical signal and realizing the operation function of different scales by selecting a specific matrix. The imaging projection module is used for diffracting matrix information carrying the optical signal into different angle sub-beams through a Dammam grating, and imaging and projecting the sub-beams on a detection camera at the rear end through a Fourier lens, so as to project the results of the accumulation of different matrix elements between diffraction orders on a detection plane. The optical signal detection module is used for recording the linear operation results of the matrix elements according to the size of the gray scale, and deducting the background noise generated by related devices, so as to realize the addition results of the gray scale of the to-be-calculated optical matrix. The addition calculation acceleration module is used for matching the size of the optical matrix element with the sliding relationship between adjacent two diffraction orders, so as to realize the addition of different size matrices and optical parallel linear operation.
2. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The light source is an LED light source, a vertical cavity surface semiconductor laser array, a digital projection processor or an optical fiber array.
3. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The light source signal loading module includes a first plano-convex lens, a 4f system, an aperture, a polarization beam splitter and a spatial light modulator; the light emitted by the light source is diffused into quasi-parallel light through the first plano-convex lens, the quasi-parallel light is further formed into a standard collimated light beam through the 4f system, the collimated light beam is reflected to the spatial light modulator through the polarization beam splitter, the spatial light modulator loads the gray scale information and converts the optical information into numerical information through coding, and the collimated light beam after modulation is reflected by the liquid crystal and changes the polarization state, and then transmits through the polarization beam splitter to complete signal loading.
4. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The 4f system includes a second plano-convex lens and a third plano-convex lens arranged oppositely, and an aperture is arranged between the second plano-convex lens and the third plano-convex lens.
5. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The imaging projection module includes a first Fourier lens, a Dammam grating and a second Fourier lens, the light signal emitted by the polarization beam splitter converges into a light spot at the focal point through the first Fourier lens, so as to facilitate the diffraction of the Dammam grating, the Dammam grating diffracts the matrix carrying the to-be-calculated information into a plurality of sub-beams containing the same light intensity information at different diffraction angles, and the sub-beams are imaged and projected on the detection plane at a specific diffraction angle through the second Fourier lens.
6. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The addition calculation acceleration module adjusts the distance between the Dammam grating and the first Fourier lens and the second Fourier lens, and the size of the matrix pixel loaded by the light source signal loading module, so that the light modulated by the Dammam grating is diffracted at a set diffraction angle, and adjacent diffraction orders slide a set distance from each other to realize the superposition of different matrix element weight values. The diffraction angle is affected by the diffraction formula as follows: ； The sliding step of the diffraction order is affected by the following formula: ； wherein , are the diffraction angles of the first and second diffraction orders of the diffraction grating, respectively, are the diffraction angles of the first and second diffraction orders of the diffraction grating, respectively, is the distance from the diffraction grating to the second Fourier lens, ; The sliding step s is matched with the matrix pixel size ∆, and the proportional relationship between s and ∆ is adjusted to realize the addition operation of different size matrices.
7. The optical high-parallel matrix addition framework based on multiple imaging projection according to claim 6, characterized in that, The Dammann grating is an even Dammann grating, for an optical matrix of size , diffracts the beam into two beams, one of which is of order (+1, +1) and the other of order (-1, -1). The matrix elements of the 11th to 20th rows of the matrix are superimposed with the elements of the 1st to 10th rows of the matrix of order (+1, -1), and the orders (+1, ) and (-1, ) are just staggered by 10 elements in the transverse diffraction direction, so as to realize the matrix addition of , and obtain the matrix .
8. The optical high-parallel matrix addition architecture based on multiple imaging projection according to claim 1, wherein, The optical signal detection module comprises a high-dynamic high-sensitivity camera and a gray scale coding unit; the camera is a CCD, CMOS or sCMOS camera, which is used to receive a specific plane gray scale matrix calculation result; the gray scale coding unit realizes natural number addition operation through gray scale quantization and assignment, and can adapt to positive and negative number and complex number addition calculation requirements.
9. The method of claim 1-8, wherein the method is a method of computing based on a multi-imaging projection-based optical high-parallel matrix addition framework, characterized in that, The method comprises the following steps: Light source signal loading: loading the matrix information to be calculated in the form of gray scale coding on the collimated light beam through a light source signal loading module to complete optical signal modulation; Multi-imaging projection: using a Dammam grating in the imaging projection module to split the light beam carrying the matrix information into multiple sub-beams containing the same information, and the sub-beams are imaged and projected to the detection plane through a Fourier lens; Addition acceleration regulation: adjusting the distance between the Dammam grating and the Fourier lens and the matrix element size through an addition calculation acceleration module to make the adjacent diffraction order sub-beams slide to a preset distance, so as to realize spatial superposition of the matrix elements; Signal detection decoding: the optical signal detection module receives the superposed gray scale matrix signal, and obtains the digital result of the matrix addition through gray scale decoding after deducting the background noise.
10. The computational method of claim 9, wherein, In the addition acceleration regulation step, the following steps are further included: The sliding distance s of the adjacent diffraction order sub-beams and the matrix element size Δ satisfy s=k×Δ, k is a positive integer, the weighted superposition of the matrix elements is realized by presetting the value of k, when k=1, the direct accumulation of the adjacent matrix elements is realized, and when k=n, the accumulation of the matrix elements with an interval of n-1 elements is realized; the sliding distance s is regulated through the distance d between the Dammam grating and the Fourier lens and the diffraction angle θ, and s=d×tanθ is satisfied.