Object recognition apparatus and object recognition method

Abstract

The present disclosure provides an object recognition device, comprising: a laser generation module, a first operation module and a second operation module. The laser generation module is configured to generate a first laser comprising MxN laser components. The first operation module is configured to receive the first laser and an object to be recognized, and perform a convolution operation on the object to be recognized by the MxN laser components to obtain convolution data. The second operation module is configured to receive the first laser and the convolution data, and perform a full connection operation on the convolution data by the MxN laser components to obtain an object recognition result. Wherein the wavelengths of the MxN laser components are different, the wavelength interval between any two adjacent laser components is equal, and M and N are positive integers. The present disclosure also provides an object recognition method.

Description

Technical Field

[0001] This disclosure relates to the fields of big data technology and microwave photonics, specifically to an object recognition device and an object recognition method. Background Technology

[0002] Neural networks, as a big data computing model with wide applications in signal processing, physics, image processing, artificial intelligence, and other fields, rely on high-performance computers to improve data processing efficiency. However, traditional electricity-based computers are facing bottlenecks in bandwidth, latency, and energy consumption. Therefore, the fact that traditional computers are gradually becoming unable to efficiently process big data has become an urgent problem to be solved. Summary of the Invention

[0003] In view of the above, in order to overcome at least one aspect of the above problems, this disclosure provides an object recognition device, comprising: a laser generating module, a first calculation module, and a second calculation module; the laser generating module is used to generate a first laser comprising M×N laser components; the first calculation module is used to receive the first laser and an object to be recognized, and to perform a convolution operation on the object to be recognized using the M×N laser components to obtain convolutional data; the second calculation module is used to receive the first laser and the convolutional data, and to perform a fully connected operation on the convolutional data using the M×N laser components to obtain an object recognition result; wherein the M×N laser components have different wavelengths, the wavelength interval between any two adjacent laser components is equal, and M and N are positive integers.

[0004] According to an embodiment of this disclosure, the first computing module includes an arbitrary waveform generator and at least one convolution operation unit; the arbitrary waveform generator is used to convert the object to be identified into a signal to be identified; the at least one convolution operation unit is used to receive the first laser and the signal to be identified, and to perform at least one convolution operation on the object to be identified through the M×N laser components to obtain convolution data; the second computing module includes at least one fully connected operation unit, the at least one fully connected operation unit is used to receive the first laser and the convolution data, and to perform at least one fully connected operation on the convolution data through the M×N laser components to obtain an object recognition result.

[0005] According to an embodiment of this disclosure, the convolution operation unit includes a first electro-optic modulator, a first dispersion medium, a first waveform shaper, a first balanced photodetector array, and a data processing unit; the first electro-optic modulator is used to receive the signal to be identified and the first laser, and modulate the signal to be identified onto the first laser; the first dispersion medium is used to receive the modulated first laser and delay the M×N laser components of the modulated first laser to obtain a second laser; the first waveform shaper is used to receive the second laser and adjust the M×N laser components of the second laser according to N preset convolution kernels. The intensity of the light is measured, and the adjusted M×N laser components are split into N sets of optical signals, each set of optical signals comprising two optical signals. The N sets of optical signals correspond to the N preset convolution kernels. The first balanced photodetector array comprises N first balanced photodetectors, which are used to receive the N sets of optical signals respectively, convert the two optical signals in each set of received optical signals into two electrical signals, and perform a difference operation on the two electrical signals to obtain multiple convolution results. The data processing unit is used to convert the multiple convolution results into one-dimensional data to obtain the convolution data.

[0006] According to an embodiment of this disclosure, the first waveform shaper includes N sets of output ports, each set of output ports including a first output port and a second output port, and each preset convolutional kernel includes M elements; the first waveform shaper is used to adjust the intensity of M×N laser components of the second laser according to the M×N elements of the N preset convolutional kernels, to obtain intensity-adjusted M×N laser components; the first waveform shaper is further used to divide the intensity-adjusted M×N laser components into N groups of optical signals, wherein the N... The optical signals in the group correspond to the N preset convolutional kernels respectively. Each group of optical signals includes a first optical signal and a second optical signal. The first waveform shaper is further used to output the first optical signal of one group of optical signals from the first output port of the corresponding group of output ports of the N groups of output ports respectively, and to output the second optical signal of one group of optical signals from the second output port of the corresponding group of output ports of the N groups of output ports respectively. Wherein, the first optical signal is the optical signal corresponding to the laser component obtained by intensity adjustment based on the negative value in the preset convolutional kernel, and the second optical signal is the optical signal corresponding to the laser component obtained by intensity adjustment based on the non-negative value in the preset convolutional kernel. The M×N elements of the N preset convolutional kernels correspond to the M×N laser components respectively.

[0007] According to an embodiment of this disclosure, the first balanced photodetector includes two input ports;

[0008] The first balanced photodetector is used to receive first and second optical signals of a set of optical signals through the two input ports respectively; the first balanced photodetector is also used to convert the first and second optical signals into first and second electrical signals respectively, and to perform a difference operation on the first and second electrical signals to obtain a convolution result.

[0009] According to an embodiment of this disclosure, the fully connected operation unit includes a first logic operation unit, which is used to receive the convolutional data and perform a fully connected operation on the convolutional data to obtain an object recognition result.

[0010] According to an embodiment of this disclosure, the fully connected computing unit includes a second electro-optic modulator, a second dispersion medium, a second waveform shaper, a second balanced photodetector array, and a second logic operation unit. The second electro-optic modulator is used to receive the convolutional data and the first laser, and modulate the convolutional data onto the first laser. The second dispersion medium is used to receive the modulated first laser and delay the M×N laser components of the modulated first laser to obtain a third laser. The second waveform shaper is used to receive the third laser, adjust the intensity of the M×N laser components of the third laser according to P preset decisions, and split the adjusted M×N laser components to obtain P groups of optical signals. Each group of optical signals includes two optical signals, and the P groups of optical signals are respectively coupled with the P... Each preset decision corresponds to one other; the second balanced photodetector array includes P second balanced photodetectors, which are used to receive P groups of optical signals respectively, convert two optical signals in each group of received optical signals into two electrical signals, and perform a difference operation on the two electrical signals respectively to obtain multiple fully connected results; the second logic operation unit is used to receive the multiple fully connected results and obtain the object recognition result by comparing the multiple fully connected results; where P is a positive integer.

[0011] According to an embodiment of this disclosure, the second waveform shaper includes P groups of output ports, each group of output ports including a third output port and a fourth output port, and each preset decision includes Q fully connected weights; the second waveform shaper is used to divide the M×N laser components of the third laser into P groups of optical signals, each group of optical signals including M×N laser components, and adjust the intensity of Q laser components in the corresponding group of optical signals according to the Q fully connected weights in each preset decision to obtain P groups of optical signals with adjusted intensity, each group of optical signals including Q laser components; the second waveform shaper is also used to output the third optical signal of one group of optical signals in the adjusted P groups of optical signals from the third output port of the corresponding group of output ports of the P groups of output ports, and output the fourth optical signal of one group of optical signals in the adjusted P groups of optical signals from the fourth output port of the corresponding group of output ports of the P groups of output ports;

[0012] The adjusted set of optical signals includes a third optical signal and a fourth optical signal. The third optical signal is the optical signal corresponding to the laser component adjusted according to the negative weight value in the preset decision. The fourth optical signal is the optical signal corresponding to the laser component adjusted according to the non-negative weight value in the preset decision. The P×Q fully connected weights of the P preset decisions correspond to P×Q laser components respectively, where Q≤M×N and Q is a positive integer.

[0013] According to an embodiment of this disclosure, the Q laser components are the Q laser components with consecutively adjacent wavelengths among the M×N laser components of the set of optical signals.

[0014] This disclosure provides an object recognition method, comprising: an object recognition device based on any one of the above-described methods, the method comprising: a laser generation module generating a first laser comprising M×N laser components; a first processing module receiving the first laser and an object to be recognized, and performing a convolution operation on the object to be recognized using the M×N laser components to obtain convolutional data; a second processing module receiving the first laser and the convolutional data, and performing a fully connected operation on the convolutional data using the M×N laser components to obtain an object recognition result; wherein the M×N laser components have different wavelengths, the wavelength interval between any two adjacent laser components is equal, and M and N are positive integers.

[0015] Compared with the prior art, this disclosure has the following beneficial effects:

[0016] 1. Utilizing the advantages of light, such as high bandwidth, high speed, and low latency, the transmission time and wavelength of light are designed to enable efficient big data processing through optical transmission.

[0017] 2. Using multiple laser components of different wavelengths to achieve parallel operation of multiple convolution kernels and multiple fully connected layers not only improves the data input rate, but also makes the data processing process faster.

[0018] 3. By using the intensity of laser components to represent the values ​​in the convolution kernel, different convolution kernels and multiple fully connected layers can be implemented simply by adjusting the number and intensity of the laser components. This gives the device high reconfigurability and scalability. Furthermore, multiple computational modules can be cascaded to achieve deeper neural network structures, demonstrating strong scalability. Attached Figure Description

[0019] To gain a more complete understanding of this disclosure and its advantages, reference will now be made to the following description taken in conjunction with the accompanying drawings, wherein:

[0020] Figure 1 A schematic diagram of an object recognition device according to an embodiment of the present disclosure is shown.

[0021] Figure 2 A schematic diagram of an object recognition device according to another embodiment of the present disclosure is shown;

[0022] Figure 3 A schematic diagram of a convolution operation unit according to an embodiment of the present disclosure is shown;

[0023] Figure 4 A schematic diagram of laser components according to an embodiment of the present disclosure is shown;

[0024] Figure 5 A schematic diagram of a fully connected computing unit according to an embodiment of the present disclosure is shown;

[0025] Figure 6 A schematic diagram of an object recognition device according to another embodiment of the present disclosure is shown; and

[0026] Figure 7 A flowchart illustrating an embodiment of the object recognition method of this disclosure is shown schematically. Detailed Implementation

[0027] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, these descriptions are merely exemplary and are not intended to limit the scope of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the following detailed description, many specific details are set forth to provide a comprehensive understanding of the embodiments of this disclosure for ease of explanation. Unless otherwise defined, the technical or scientific terms used in this disclosure should have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] Figure 1 A schematic diagram of an object recognition device according to an embodiment of the present disclosure is shown. Figure 1 As shown, this disclosure provides an object recognition device 100, including: a laser generating module 1, a first calculation module 2, and a second calculation module 3.

[0030] Laser generating module 1 and first computing module 2 are connected via fiber optic patch cords, and laser generating module 1 and second computing module 3 are connected via fiber optic patch cords. First computing module 2 and second computing module 3 are connected via cables.

[0031] Laser generation module 1 is used to generate a first laser comprising M×N laser components.

[0032] For example, the laser generating module 1 can be a multi-wavelength laser. The laser generating module 1 may also include a laser array and a beam combining module. The laser array is used to generate M×N lasers. The beam combining module is used to receive and couple the M×N lasers to obtain a first laser. The beam combining module includes, but is not limited to, an optical coupler, an arrayed waveguide grating, a wavelength division multiplexer, and a dense wavelength division multiplexer.

[0033] Laser generating module 1 generates a laser with a comb-shaped spectrum, which has M×N laser components. Each laser component has a different wavelength, and the wavelength interval between any two adjacent laser components is equal. M and N are positive integers.

[0034] The first processing module 2 is used to receive the first laser and the object to be identified, and to perform convolution operation on the object to be identified using M×N laser components to obtain convolution data.

[0035] The second processing module 3 is used to receive the first laser and convolutional data, and to perform a fully connected operation on the convolutional data through M×N laser components to obtain the object recognition result.

[0036] The objects to be identified include, but are not limited to, images, audio, video, and text. For example, if the object to be identified is an image, convolution and fully connected operations can be performed on complex images to identify text and images within them.

[0037] In this embodiment of the disclosure, multiple laser components are used as carriers to realize multi-level convolution operations and fully connected operations. Taking advantage of the advantages of light, such as large bandwidth, high speed and low latency, the transmission time and wavelength of light are designed to achieve efficient big data processing through optical transmission.

[0038] Figure 2 A schematic diagram of an object recognition device according to another embodiment of the present disclosure is shown. Figure 2 As shown, the object recognition device 200 includes: a laser generation module 1, a first calculation module 2, and a second calculation module 3. The first calculation module 2 includes an arbitrary waveform generator 21 and at least one convolution operation unit 22. The second calculation module 3 includes at least one fully connected operation unit 31.

[0039] The arbitrary waveform generator 21 is connected to the convolution operation unit 22 via a cable, and multiple convolution operation units 22 and multiple fully connected operation units 31 are sequentially connected via cables. The laser generation module 1 is connected to multiple convolution operation units 22 and multiple fully connected operation units 31 via fiber optic patch cords.

[0040] Arbitrary waveform generator 21 converts the object to be identified into a signal to be identified.

[0041] The objects to be identified include, but are not limited to, images, audio, video, and text. An arbitrary waveform generator converts the images, audio, and text to be processed into electrical signals to be identified.

[0042] At least one convolution operation unit 22 receives a first laser and a signal to be identified, and performs at least one convolution operation on the object to be identified using M×N laser components to obtain convolution data.

[0043] At least one fully connected operation unit 31 receives the first laser and convolutional data, and performs at least one fully connected operation on the convolutional data through M×N laser components to obtain the object recognition result.

[0044] In this embodiment of the disclosure, multiple convolutional operation units and multiple fully connected operation units are set to construct a deeper neural network structure. By performing multiple convolutional and fully connected operations on the object to be identified, more refined object recognition can be achieved, and the accuracy of object recognition can be improved.

[0045] Figure 3 A schematic diagram of a convolution operation unit according to an embodiment of the present disclosure is shown. Figure 3 As shown, the convolution operation unit 22 includes a first electro-optic modulator 221, a first dispersive medium 222, a first waveform shaper 223, a first balanced photodetector array 224, and a data processing unit 225.

[0046] The first electro-optic modulator 221, the first dispersive medium 222, the first waveform shaper 223, and the first balanced photodetector array 224 are connected in sequence via fiber optic patch cords. The first balanced photodetector array 224 is connected to the data processing unit 225 via a cable.

[0047] The first electro-optic modulator 221 is used to receive the signal to be identified and the first laser, and modulate the signal to be identified onto the first laser.

[0048] The signal to be identified is used as a modulation signal and modulated onto the first laser by the first electro-optic modulator 221. Understandably, the object to be identified is loaded onto the first laser by the first electro-optic modulator 221 to change the intensity of the first laser.

[0049] It should be noted that the intensity of the first laser beam changes after modulation. The intensity of all M×N laser components in the modulated first laser beam changes, and the intensity of each laser component changes proportionally. The intensity relationship between the modulated laser components is consistent with the intensity relationship between the laser components before modulation.

[0050] After receiving the modulated first laser, the first dispersive medium 222 delays the M×N laser components of the modulated first laser to obtain the second laser. The first dispersive medium 222 delays the M×N laser components to different degrees.

[0051] The first dispersion medium 222 includes, but is not limited to, at least one of dispersion-compensating fiber, chirped fiber grating, ordinary single-mode fiber, and multimode fiber.

[0052] The first dispersive medium 222 can also be replaced by a combination of a first arrayed waveguide grating, a second arrayed waveguide grating, and an optical delay line array. The first arrayed waveguide grating, the optical delay line array, and the second arrayed waveguide grating are connected sequentially. The first arrayed waveguide grating can also be replaced by a wavelength division multiplexer, a dense wavelength division multiplexer, or other devices that can split different wavelengths into a single beam. The second arrayed waveguide grating can be replaced by a wavelength division multiplexer, a dense wavelength division multiplexer, an optical coupler, or other devices that can combine different light wavelengths into a single beam.

[0053] Figure 4 A schematic diagram of laser components according to an embodiment of the present disclosure is shown.

[0054] like Figure 4As shown in (A), the laser generating module 1 generates a first laser L, which includes four laser components L1 to L4. The wavelengths of the laser components L1 to L4 are λ1, λ2, λ3, and λ4, respectively, and the wavelength relationship satisfies λ1-λ2=λ2-λ3=λ3-λ4. The first electro-optic modulator 221 modulates the signal to be identified onto the first laser L to obtain the modulated first laser L. X .like Figure 4 As shown in (B), the first dispersive medium 222 modulates the first laser L. X The four laser components are delayed to obtain the second laser L. Y .

[0055] In this process, the signal to be identified is periodically modulated onto the first laser L, and the first laser L is changed in each cycle (i.e. Figure 4 Each signal X in (A) i The intensity within a duration of time. This period can be considered the signal input period. For example, such as Figure 4 As shown in (A), the modulated first laser L X Seven signals, X1 to X7, are loaded onto the device. For example... Figure 4 As shown in (B), the first dispersive medium 222 modulates the first laser L. X The four laser components are delayed such that there is a delay of one signal input cycle between any two laser components with adjacent wavelengths. For example, after group velocity dispersion by the first dispersive medium 222, the propagation time of laser component L2 is delayed by one input cycle relative to laser component L1, the propagation time of laser component L3 is delayed by one input cycle relative to laser component L2, and the propagation time of laser component L4 is delayed by one input cycle relative to laser component L3.

[0056] This disclosure provides an exemplary method for achieving a delay of one signal input period between any two laser components with adjacent wavelengths. However, this disclosure does not limit the specific laser delay method.

[0057] For example, the laser component and the first dispersive medium 222 can be configured to satisfy the following:

[0058]

[0059] Where Δλ is the wavelength interval between any two adjacent laser components, f k The input frequency of the signal to be identified. Let L be the dispersion coefficient of the first dispersive medium 222 and L be the length of the first dispersive medium 222. Understandably, the delay between any two laser components with adjacent wavelengths is 1 / f. k .

[0060] For example, Figure 4 (A) shows the first laser L after being modulated by the signal to be identified. X The four laser components. Figure 4 (B) The second laser L Y The signal to be identified, Y1 to Y2, is loaded onto the top. 10 It is determined by the signals X1 to X7 to be identified loaded onto the laser components L1 to L4 before dispersion. For example:

[0061] Y1 = X1

[0062] Y2 = X2 + X1

[0063] Y3 = X3 + X2 + X1

[0064] Y4 = X4 + X3 + X2 + X1

[0065] Y5~Y 10 The expression can be found in the reference. Figure 4 (B) and the expressions for Y1 to Y4 mentioned above will not be repeated in this disclosure.

[0066] See also Figure 3 The first waveform shaper 223 receives the second laser, adjusts the intensity of the M×N laser components of the second laser according to N preset convolution kernels, and splits the adjusted M×N laser components to obtain N sets of optical signals.

[0067] The first waveform shaper 223 can also be replaced by a wavelength selection switch. The first waveform shaper 223 attenuates the intensity of different laser components so that the relative intensity relationship of each laser component after attenuation is the same as the relative relationship of the convolution kernel midpoint.

[0068] Among them, N sets of optical signals correspond to N preset convolution kernels, each preset convolution kernel includes M elements, and the M×N elements of the N preset convolution kernels correspond to M×N laser components.

[0069] For example, a preset convolution kernel can be a matrix A, which contains M elements. The M×N elements of the N preset convolution kernels correspond to the M×N laser components.

[0070] Quote Figure 1 The illustrated embodiment is an example of this. Laser generating module 1 generates 4×3 laser components L1 to L2. 12 The intensity of 4×3 laser components is adjusted according to 3 preset convolution kernels, each preset convolution kernel including 4 elements. Among them, the first waveform shaper 223 adjusts the intensity of one laser component according to one element, and every 4 laser components correspond to the 4 elements of one preset convolution kernel.

[0071] The first waveform shaper 223 includes N sets of output ports, each set of output ports including a first output port and a second output port.

[0072] The first waveform shaper 223 outputs the first optical signal of one group of optical signals from the first output port of the corresponding N groups of output ports, and outputs the second optical signal of one group of optical signals from the second output port of the corresponding N groups of output ports.

[0073] The optical signal set includes a first optical signal and a second optical signal. The first optical signal is the optical signal corresponding to the laser component obtained by intensity adjustment based on the negative values ​​in the preset convolution kernel, and the second optical signal is the optical signal corresponding to the laser component obtained by intensity adjustment based on the non-negative values ​​in the preset convolution kernel.

[0074] The first balanced photodetector array 224 includes N first balanced photodetectors, which are used to receive N sets of optical signals respectively.

[0075] The first balanced photodetector receives one set of optical signals from the N sets of optical signals obtained by the first waveform shaper 223. The first balanced photodetector includes two input ports. After receiving the first and second optical signals from the first set of optical signals through the two input ports, the first balanced photodetector converts the first and second optical signals into first and second electrical signals respectively, and performs a difference operation on the first and second electrical signals to obtain a convolution result. N convolution results can be obtained through N first balanced photodetectors.

[0076] For example, quoting again Figure 4 In the listed embodiments, the first waveform shaper 223 adjusts the intensity of four laser components L1 to L4 according to four elements W1 to W4 of a preset convolution kernel. Elements W2 and W4 are negative, while elements W1 and W3 are positive. The laser components L2 and L4 adjusted by elements W2 and W4 form the first optical signal, which is output from the first output port of a set of optical signals. The laser components L1 and L3 adjusted by elements W1 and W3 form the second optical signal, which is output from the second output port of the same set of optical signals.

[0077] The intensity-adjusted signal Y' to be identified can be represented as:

[0078] Y'1=X1|W1|

[0079] Y'2=X2|W1|+X1|W2|

[0080] Y'3 = X3|W1| + X2|W2| + X1|W3|

[0081] Y'4=X4|W1|+X3|W2|+X2|W3|+X1|W4|

[0082] Y'5~Y' 10 The expression can be found in the reference. Figure 4 (B) and the expressions for Y'1 to Y'4 above will not be repeated in this disclosure. The elements W of the preset convolution kernel include negative values.

[0083] A first balanced photodetector receives a set of optical signals output from the same set of output ports through two input ports. The first optical signal, composed of laser components L2 and L4, is input through one of the two input ports, while the second optical signal, composed of laser components L1 and L3, is input through the other input port. The first balanced photodetector also converts the received first and second optical signals into first and second electrical signals, respectively. The difference between the first and second electrical signals is then calculated. The electrical signals output by the first balanced photodetector are y'1~y' 10 ,y'1~y' 10 Respectively with Y'5~Y' 10 Corresponding. For example,

[0084] y'1=α(X1|W1|)

[0085] y'2=α(X2|W1|-X1|W2|)

[0086] y'3=α(X3|W1|-X2|W2|+X1|W3|)

[0087] y'4=α(X4|W1|-X3|W2|+X2|W3|-X1|W4|)

[0088] Where α is the response of the first balanced photodetector. y'5~y' 10 The expressions can be found in the above y'5~y'. 10 The expression for this will not be elaborated upon in this disclosure.

[0089] The waveform shaper adjusts the intensity of the laser component according to the preset convolution and distinguishes the light signal adjusted according to different preset convolution kernels based on wavelength selection. The detector array performs subtraction on the light signal adjusted according to different preset convolution kernels. Thus, by combining multiple wavelengths with multiple elements of the convolution kernel, the operation of multiple convolution kernels containing multiple elements can be realized.

[0090] The data processing unit 225 is used to convert multiple convolution results into one-dimensional data to obtain convolution data.

[0091] The convolution result is usually an analog signal. The data processing unit 225 converts the multiple parallel convolution results from analog signals into digital signals, and then converts the multiple digital signals into one-dimensional data, realizing the conversion of the data format from parallel to serial, so that the second operation module 3 can perform fully connected operations on the operation results of the first operation module 2.

[0092] Before the data processing unit 225 converts multiple convolution results into one-dimensional convolution data, it can also pool multiple parallel convolution results to reduce the amount of data, thereby reducing the complexity of subsequent fully connected operations.

[0093] This disclosure utilizes laser light as a carrier to perform convolution operations on data, leveraging the advantages of light such as high bandwidth, high speed, and low latency to improve computational efficiency. Parallel convolution operations with multiple kernels using multiple laser components of different wavelengths not only increase the data input rate but also make the data processing faster. Furthermore, the object recognition device of this disclosure, without changing the device hardware, only requires adjusting the number and intensity of the laser components to adjust the number and dimensions of the convolution kernels involved in the operation, thereby achieving the ability to perform convolution operations with different numbers and dimensions. This gives the device high reconfigurability and scalability.

[0094] Figure 5 A schematic diagram of a fully connected computing unit according to an embodiment of the present disclosure is shown. Figure 5 As shown, the fully connected operation unit 31 includes a second electro-optic modulator 311, a second dispersion medium 312, a second waveform shaper 313, a second balanced photodetector array 314, and a second logic operation unit 315.

[0095] The second electro-optic modulator 311, the second dispersive medium 312, the second waveform shaper 313, and the second balanced photodetector array 314 are connected sequentially via fiber optic patch cords. The second balanced photodetector array 314 and the second logic operation unit 315 are connected via cables.

[0096] The second electro-optic modulator 311 receives convolutional data and the first laser, and modulates the convolutional data onto the first laser.

[0097] It should be noted that the convolution data output by the first arithmetic module 2 is a digital signal, which needs to be converted from a digital signal to an analog signal by a digital-to-analog converter. The second electro-optic modulator 311 modulates the convolution data onto the first laser.

[0098] The second dispersive medium 312 receives the modulated first laser and delays the M×N laser components of the modulated first laser to obtain the third laser.

[0099] The specific implementation details and technical effects of the second electro-optic modulator 311 are the same as those of the first electro-optic modulator 221, and the specific implementation details and technical effects of the second dispersive medium 312 are also the same as those of the first dispersive medium 222. These will not be repeated in the embodiments disclosed herein.

[0100] After receiving the third laser, the second waveform shaper 313 adjusts the intensity of the M×N laser components of the third laser according to P preset decisions, and then splits the adjusted M×N laser components into P groups of optical signals. Each group of optical signals includes two optical signals, and the P groups of optical signals correspond one-to-one with the P preset decisions. Each preset decision includes Q fully connected weights, where Q and P are positive integers.

[0101] The second waveform shaper 313 can also be replaced by a wavelength selection switch. The second waveform shaper 313 attenuates the intensity of different laser components so that the relative intensity relationship of each laser component after attenuation is the same as the relative relationship of the preset decision median value.

[0102] The second waveform shaper 313 performs power division on the M×N laser components based on the number of fully connected weights (P×Q) included in the P preset decisions. If the number of laser components obtained after power division is greater than the number of fully connected weights, the second waveform shaper 313 can filter out the excess laser components.

[0103] For example, the second waveform shaper 313 divides the M×N laser components of the third laser into P groups of optical signals, each group containing M×N laser components, and each preset decision corresponds to one of the groups of optical signals in the P groups.

[0104] Based on the Q fully connected weights in each preset decision, the intensity of Q laser components in a set of M×N laser components of an optical signal is adjusted to obtain P sets of intensity-adjusted optical signals, each set of optical signals including Q laser components. Specifically, the Q laser components in each set of optical signals are the Q laser components with consecutively adjacent wavelengths from the original M×N laser components of that set of optical signals, where Q ≤ M×N.

[0105] The second waveform shaper 313 includes P groups of output ports, each group of output ports including a third output port and a fourth output port.

[0106] The second waveform shaper 313 splits the intensity-adjusted P×Q laser components into P groups of optical signals, and outputs the third optical signal of one group of optical signals from the third output port of the corresponding P group output port, and outputs the fourth optical signal of one group of optical signals from the fourth output port of the corresponding P group output port.

[0107] The adjusted set of optical signals includes a third optical signal and a fourth optical signal. The third optical signal is the optical signal corresponding to the laser component adjusted according to the negative weight value in the preset decision, and the fourth optical signal is the optical signal corresponding to the laser component adjusted according to the non-negative weight value in the preset decision.

[0108] The second balanced photodetector array 314 includes P second balanced photodetectors, which are used to receive P groups of optical signals respectively.

[0109] The second balanced photodetector receives one set of optical signals from the P sets of optical signals obtained by the second waveform shaper 313. The second balanced photodetector includes two input ports. After receiving the third and fourth optical signals from the set of optical signals through the two input ports, the second balanced photodetector converts the third and fourth optical signals into third and fourth electrical signals, and performs a difference operation on the third and fourth electrical signals respectively to obtain multiple fully connected results. P fully connected results can be obtained through P second balanced photodetectors.

[0110] The specific implementation details and technical effects of the second balanced photodetector array 314 are the same as those of the first balanced photodetector array 214, and will not be repeated in this disclosure.

[0111] The second logic operation unit 315 receives multiple fully connected results and obtains the object recognition result by comparing the multiple fully connected results.

[0112] The fully connected result output by the second balanced photodetector array 314 is an analog signal. The second logic operation unit 315 converts the fully connected result from the analog signal into a digital signal and compares multiple fully connected results in digital signal form to obtain the object recognition result.

[0113] This disclosure utilizes laser light as a carrier to perform fully connected computations on data, leveraging the advantages of light such as high bandwidth, high speed, and low latency to improve computational efficiency. Parallel fully connected computations using multiple laser components with different wavelengths for multiple fully connected decisions not only increase the data input rate but also make the data processing faster. Furthermore, the object recognition device of this disclosure, without changing the device hardware, only needs to adjust the number and intensity of the laser components to adjust the number and dimensions of decisions, thereby achieving computational capabilities for different numbers and dimensions of decisions. This gives the device high reconfigurability and scalability.

[0114] It should be noted that the number and dimensions of convolutional kernels and fully connected decision pairs listed in this disclosure are merely illustrative. Those skilled in the art can set the appropriate number of convolutional kernels and fully connected decision pairs according to actual computational needs. This disclosure does not limit the number and dimensions of convolutional kernels in convolutional operations, nor does it limit the number and dimensions of fully connected decision pairs in fully connected operations. The number and dimensions of convolutional kernels and fully connected decision pairs are typically determined through pre-training.

[0115] Figure 6 A schematic diagram of an object recognition device according to another embodiment of the present disclosure is shown. Figure 6 As shown, the fully connected arithmetic unit 31 includes a first logic unit 316. The first logic unit 316 is connected to the first arithmetic unit 2 via a cable.

[0116] The first logic operation unit 316 receives convolutional data and performs a fully connected operation on the convolutional data output by the first operation unit 2 to obtain the object recognition result.

[0117] The convolutional data includes multiple parallel convolutional results. Before performing fully connected operations on the convolutional data, the first logic operation unit 316 needs to convert the multiple convolutional results from analog signal form to digital signal form, and pool the multiple digital signals to reduce the computational load of subsequent fully connected operations. Then, the pooled multiple digital signals are converted into one-dimensional data, realizing the conversion of the data format from parallel to serial. Finally, fully connected calculations are performed directly on the serial data to obtain the object recognition result.

[0118] As an optional embodiment, the object recognition device 100 may further include an electrical amplifier and an optical amplifier in addition to the object recognition device 100.

[0119] An electrical amplifier can be located anywhere in the electrical link to amplify the power of the electrical signal and send the amplified signal to be processed to the electro-optic modulator. An optical amplifier can be located anywhere in the optical link to amplify the power of the laser in the optical link.

[0120] This disclosure provides a detailed object recognition method applicable to the aforementioned object recognition device. Figure 7 The illustration is a flowchart of an object identification method according to an embodiment of the present disclosure.

[0121] like Figure 7 As shown, the object recognition method includes at least the following steps:

[0122] S1, the laser generation module generates a first laser consisting of M×N laser components.

[0123] S2, the first processing module receives the first laser and the object to be identified, and performs convolution operation on the object to be identified using M×N laser components to obtain convolution data.

[0124] S3, the second processing module receives the first laser and convolutional data, and performs a fully connected operation on the convolutional data using M×N laser components to obtain the object recognition result.

[0125] Among them, the wavelengths of the M×N laser components are different, and the wavelength interval between any two adjacent laser components is equal, where M and N are positive integers.

[0126] It should be noted that the object recognition method in the embodiments of this disclosure corresponds to the object recognition device in the embodiments of this disclosure. For a detailed description of the object recognition method, please refer to the object recognition device section, which will not be repeated here.

[0127] It should also be noted that, without conflict, the embodiments and features of the embodiments of this disclosure can be combined with each other to obtain new embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and are not intended to limit it. Although this disclosure has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of this disclosure without departing from the spirit and scope of the technical solutions of this disclosure.

Claims

1. An object recognition apparatus characterized by comprising: The method comprises the following steps: A laser generation module, a first operation module and a second operation module are included; The laser generation module is configured to generate a first laser comprising M×N laser components; The first operation module is configured to receive the first laser and an object to be identified, and perform convolution operation on the object to be identified by using the M×N laser components to obtain convolution data; The second operation module is configured to receive the first laser and the convolution data, and perform full connection operation on the convolution data by using the M×N laser components to obtain an object recognition result; Wherein, the M×N laser components have different wavelengths, and the wavelength interval between any two adjacent laser components is equal, and M and N are positive integers; The first operation module comprises an arbitrary waveform generator and at least one convolution operation unit; the arbitrary waveform generator is configured to convert the object to be identified into an identification signal; and the at least one convolution operation unit is configured to receive the first laser and the identification signal, and perform at least one convolution operation on the object to be identified by using the M×N laser components to obtain the convolution data; The second operation module comprises at least one full connection operation unit, which is configured to receive the first laser and the convolution data, and perform at least one full connection operation on the convolution data by using the M×N laser components to obtain the object recognition result; The convolution operation unit comprises a first electro-optical modulator, a first dispersive medium, a first waveform shaper, a first balanced photodetector array and a data processing unit; The first electro-optical modulator is configured to receive the identification signal and the first laser, and modulate the identification signal onto the first laser; The first dispersive medium is configured to receive the modulated first laser, and delay the M×N laser components of the modulated first laser to obtain a second laser; The first waveform shaper is configured to receive the second laser, adjust the intensity of the M×N laser components of the second laser according to N preset convolution kernels, and split the adjusted M×N laser components to obtain N groups of optical signals, wherein each group of optical signals comprises two optical signals; The first balanced photodetector array comprises N first balanced photodetectors, which are configured to respectively receive N groups of optical signals, respectively convert the two optical signals in each group of received optical signals into two electrical signals, and respectively perform difference operation on the two electrical signals to obtain a plurality of convolution results; The data processing unit is configured to convert the plurality of convolution results into one-dimensional data to obtain the convolution data; Wherein, the N groups of optical signals correspond to the N preset convolution kernels respectively; each of the preset convolution kernels comprises M elements, and the M×N elements of the N preset convolution kernels correspond to the M×N laser components respectively.

2. The apparatus of claim 1, wherein, The first waveform shaper comprises N output ports, and each output port comprises a first output port and a second output port; The first waveform shaper is further configured to output first optical signals of a group of optical signals from a first output port of a group of output ports corresponding to the first optical signals respectively, and output second optical signals of the group of optical signals from a second output port of the group of output ports corresponding to the second optical signals respectively. The first optical signals are optical signals corresponding to laser components obtained by adjusting intensities according to negative values in a preset convolution kernel, and the second optical signals are optical signals corresponding to laser components obtained by adjusting intensities according to non-negative values in the preset convolution kernel.

3. The apparatus of claim 1 or 2, wherein, The first balanced photodetector includes two input ports. The first balanced photodetector is configured to receive first and second optical signals of a group of optical signals through the two input ports respectively. The first balanced photodetector is further configured to convert the first and second optical signals into first and second electrical signals respectively, and perform a difference operation on the first and second electrical signals to obtain a convolution result.

4. The apparatus of claim 1, wherein, The fully-connected operation unit includes a first logical operation unit configured to receive the convolution data and perform a fully-connected operation on the convolution data to obtain an object recognition result.

5. The apparatus of claim 1, wherein, The fully-connected operation unit includes a second electro-optical modulator, a second dispersive medium, a second waveform shaper, a second balanced photodetector array, and a second logical operation unit. The second electro-optical modulator is configured to receive the convolution data and the first laser, and modulate the convolution data onto the first laser. The second dispersive medium is configured to receive the modulated first laser, and delay M×N laser components of the modulated first laser to obtain a third laser. The second waveform shaper is configured to receive the third laser, adjust intensities of the M×N laser components of the third laser according to P preset decisions, and split the adjusted M×N laser components to obtain P groups of optical signals, one group of optical signals including two optical signals, the P groups of optical signals corresponding to the P preset decisions one by one respectively. The second balanced photodetector array includes P second balanced photodetectors, the P second balanced photodetectors being configured to receive P groups of optical signals respectively, convert two optical signals in a group of optical signals received by each second balanced photodetector into two electrical signals respectively, and perform a difference operation on the two electrical signals respectively to obtain a plurality of fully-connected results. The second logical operation unit is configured to receive the plurality of fully-connected results, and obtain an object recognition result by comparing the plurality of fully-connected results. P is a positive integer.

6. The apparatus of claim 5, wherein, The second waveform shaper includes P groups of output ports, each group of output ports including a third output port and a fourth output port, and each preset decision includes Q fully-connected weights. The second waveform shaper is configured to split the M*N laser components of the third laser into P groups of optical signals, each group of optical signals including M*N laser components, adjust the intensity of Q laser components in the M*N laser components of each group of optical signals according to Q full-connection weights in each preset decision, and obtain P groups of optical signals with adjusted intensity, each group of optical signals including Q laser components. The second waveform shaper is further configured to output a third optical signal of one group of optical signals from a third output port of one group of output ports corresponding to the P groups of output ports, and output a fourth optical signal of the one group of optical signals from a fourth output port of the one group of output ports. The one group of optical signals with adjusted intensity includes the third optical signal and the fourth optical signal, the third optical signal being an optical signal corresponding to a laser component adjusted according to a negative weight value in a preset decision, and the fourth optical signal being an optical signal corresponding to a laser component adjusted according to a non-negative weight value in the preset decision, P*Q full-connection weights of the P preset decisions corresponding to P*Q laser components, and Q being a positive integer and less than M*N.

7. The apparatus of claim 6, wherein, The Q laser components are Q laser components with continuous and adjacent wavelengths in the M*N laser components of the one group of optical signals.

8. An object recognition method characterized by, The object recognition device based on any one of claims 1-7, the method comprising: a laser generation module generating a first laser including M*N laser components; a first operation module receiving the first laser and an object to be identified, and performing convolution operation on the object to be identified through the M*N laser components to obtain convolution data; a second operation module receiving the first laser and the convolution data, and performing full-connection operation on the convolution data through the M*N laser components to obtain an object recognition result; wherein the M*N laser components have different wavelengths, the wavelength interval between any two laser components with adjacent wavelengths is equal, and M and N are positive integers.

Images