Portable diffraction neural network image recognition system

Abstract

A portable diffraction neural network image recognition system disclosed by the present invention comprises a light input module, a light processing module, a light calculation module and a light detection module, the light input module is connected with the light calculation module through the light processing module, and the light calculation module is connected with the light detection module through the light processing module; wherein the light input module is used for acquiring a spectral image light field, the light processing module is used for processing the acquired spectral image light field, the light calculation module is used for performing diffraction neural network calculation on the light field, and the light detection module is used for classifying calculation results so as to output an image recognition result. According to the scheme, by optimizing the light path design and the 3D printing integrated structure and through input image array expansion processing, the inter-layer diffraction distance is remarkably shortened, and the system size is greatly reduced while the recognition accuracy is guaranteed to be 90% or above.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, and specifically relates to a portable diffraction neural network image recognition system. Background Technology

[0002] Image recognition technology is an important research direction in artificial intelligence, and it is widely used in navigation, map registration, natural resource analysis, environmental monitoring, medical diagnosis, and many other fields. However, traditional image recognition methods based on electronic computing suffer from high computational complexity, large data volume, high consumption of computing resources, and high processing latency.

[0003] Optical neural networks (ONNs), as an emerging computing architecture, can leverage the parallelism and high-speed propagation of photons to perform intensive operations such as matrix multiplication at the speed of light, significantly reducing energy consumption and latency, and showing potential to replace traditional artificial neural networks. Among them, diffractive neural networks (DPNNs) achieve forward computation of the neural network through the diffraction and interference of light, possessing inherent all-optical parallel processing capabilities, eliminating the need for frequent photoelectric conversion, and further improving energy efficiency and speed.

[0004] Existing diffraction neural network systems typically employ reflection or folding optical path designs, resulting in complex system structures, difficult assembly, and additional energy loss and alignment errors due to multiple reflections, affecting the overall signal-to-noise ratio and recognition stability. In contrast, this invention uses a linear optical axis transmission structure, avoiding complex reflective elements, simplifying the optical path layout, and improving system reliability and imaging quality.

[0005] Existing systems have fixed structures, making it difficult to flexibly adjust the number of network layers and diffraction distance, thus limiting their adaptability to tasks of varying complexity. For example, single-layer network structures have limited recognition capabilities when processing complex images, while multi-layer networks are difficult to integrate due to size limitations. This invention, through a replaceable SLM base and a sliding adjustment mechanism, achieves flexible configuration of the number of network layers and diffraction distance, balancing efficiency for simple tasks with the high precision requirements of complex tasks.

[0006] Therefore, how to achieve miniaturization and portability of diffraction neural network systems while maintaining high recognition accuracy, and improve their adaptability to practical image tasks, has become a key challenge for the practical application of this technology. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a portable diffraction neural network image recognition system. By optimizing the optical path design and mechanical structure, a compact, highly accurate, and portable all-optical neural network computing device is achieved, solving the technical challenges of the large size and difficulty in on-site deployment of traditional optical neural network systems.

[0008] The specific technical solution for achieving the objective of this invention is as follows:

[0009] A portable diffraction neural network image recognition system includes an optical input module, an optical processing module, an optical computing module, and an optical detection module;

[0010] The optical input module is connected to the optical processing module and the optical computing module, and the optical computing module is connected to the optical processing module and the optical detection module.

[0011] The light input module is used to acquire the light field of the spectral image, the light processing module is used to process the acquired light field of the spectral image, the light calculation module is used to perform diffraction neural network calculation on the light field, and the light detection module is used to classify the calculation results, thereby outputting the image recognition result.

[0012] Furthermore, the light input module uses optical elements to obtain a spectral image light field with a certain light intensity distribution. The light input module is composed of multiple array units and adopts an OLED screen or an LED array panel.

[0013] Furthermore, the light processing module includes a first processing unit and a second processing unit, which are used to filter out ambient stray light and make the light beam in the optical path linearly polarized.

[0014] The first processing unit is disposed as a polarizer between the optical input module and the optical computing module, and the second processing unit is disposed as an analyzer between the optical processing module and the optical detection module.

[0015] The polarization directions of the first processing unit and the second processing unit are perpendicular.

[0016] Furthermore, the first processing unit and the second processing unit are polarizers, polarizing prisms, or waveplates.

[0017] Furthermore, the optical computing module includes one or more optical modulation units;

[0018] The optical modulation unit employs one of the following: a spatial light modulator, a mask, or a metasurface.

[0019] The optical modulation unit receives the light field processed by the optical processing module, modulates the amplitude information of the incident light field, realizes the multiplication and addition calculation in the diffraction neural network, completes the diffraction propagation and interference of the light field, and outputs the calculated light field distribution.

[0020] Furthermore, the spatial light modulator is a liquid crystal spatial light modulator with a pixel size of 36 μm and a resolution of not less than 380×380. It operates in amplitude modulation mode and achieves precise control of the incident light field by loading pre-trained parameters for each pixel unit.

[0021] Furthermore, the optical detection module is a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) sensor, used to receive the output light field of the optical computing module, convert the optical signal into an electrical signal, use a pre-trained neural network to make classification decisions, and finally output the recognition result.

[0022] Furthermore, the system also includes a housing module;

[0023] The front end of the outer casing module is provided with a structure a for placing the optical input module;

[0024] The outer casing module is provided with a base for fixing the optical modulation unit and the optical detection module;

[0025] A groove is provided between the base of the optical modulation unit and the photodetector module to adjust the distance between them.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] (1) Compared with traditional optical neural network systems that require a fixed optical platform and are bulky, this invention significantly reduces the system size by optimizing the optical path design and 3D printing integrated structure. The overall size is only 372×100×154mm, achieving truly portable deployment.

[0028] (2) Compared with the traditional diffraction neural network which requires a long diffraction distance, the present invention significantly shortens the interlayer diffraction distance by arraying and expanding the input image. The interlayer spacing in the double-layer network structure is only 4.5cm, which greatly reduces the system size while ensuring a recognition accuracy of more than 90%.

[0029] (3) The present invention adopts a modular 3D printing structure design. Through precision screw hole positioning and sliding groove mechanism, it realizes the precise alignment and spacing adjustment of optical elements, solves the problem of difficult element alignment in free space optical system, and improves system stability and repeatability;

[0030] (4) The optical path design of the present invention adopts a straight transmission structure, and the optical field propagates along the same axis from the input end to the output end, avoiding complex reflection and deflection optical paths, simplifying the system structure, reducing assembly difficulty, reducing optical energy loss, and improving the system signal-to-noise ratio.

[0031] (5) The present invention supports two computing architectures, single-layer network and dual-layer network, through the design of a replaceable SLM base. It can be flexibly configured according to the complexity of the task, which not only meets the low power consumption requirements of simple tasks, but also ensures the high accuracy requirements of complex tasks, thus achieving an optimized balance between computing performance and energy consumption.

[0032] The present invention will be further described below with reference to specific embodiments. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the optical path of the portable diffraction neural network image recognition system of the present invention.

[0034] Figure 2 This is a physical image of the portable diffraction neural network image recognition system according to an embodiment of the present invention.

[0035] Figure 3 This is a schematic diagram of the arrangement of detection units in the optical detection module in an embodiment of the present invention.

[0036] Figure 4 This is a schematic diagram of the outer casing module of the portable diffraction neural network image recognition system of the present invention.

[0037] Figure 5 This is a schematic diagram of the component structure of the optical input module placed at the front end of the outer shell module in an embodiment of the present invention.

[0038] Figure 6 This is a schematic diagram of the two fixing plates on both sides of the outer shell module in an embodiment of the present invention.

[0039] Figure 7 This is a schematic diagram of the base plate structure of the outer shell module in an embodiment of the present invention.

[0040] Figure 8 This is a schematic diagram of the top plate and cable outlet structure of the outer casing module in an embodiment of the present invention.

[0041] Figure 9 This is a schematic diagram of the base structure used to fix the SLM in the outer shell module in an embodiment of the present invention.

[0042] Figure 10 This is a schematic diagram of the camera used to fix the light detection module in the outer shell module of an embodiment of the present invention. Detailed Implementation

[0043] Example

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0046] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0047] Combination Figure 1 and Figure 2 A portable diffraction neural network image recognition system includes an optical input module 1, an optical processing module 2, an optical computing module 3, and an optical detection module 4.

[0048] The optical input module 1 is connected to the optical processing module 2 and the optical computing module 3, and the optical computing module 3 is connected to the optical processing module 2 and the optical detection module 4.

[0049] The light input module 1 is used to acquire the light field of the spectral image, the light processing module 2 is used to process the acquired light field of the spectral image, the light calculation module 3 is used to perform diffraction neural network calculation on the light field, and the light detection module 4 is used to classify the calculation results, thereby outputting the image recognition results.

[0050] The light input module 1 uses optical elements to obtain a spectral image light field with a certain light intensity distribution, which can display a pre-processed extended handwritten digital image. The light input module 1 is composed of multiple array units, using an OLED screen or an LED array panel.

[0051] The emitted light from the light input module 1 is directly input into the light processing module 2 in free space. The light processing module 2 is used to process the light field of the acquired spectral image, including:

[0052] By filtering out stray light from the environment and polarizing the light field, linearly polarized light is obtained.

[0053] Specifically, the light processing module 2 includes a first processing unit and a second processing unit, which are used to filter out ambient stray light and make the light beam in the optical path linearly polarized.

[0054] The first processing unit is disposed as a polarizer between the optical input module 1 and the optical computing module 3, and the second processing unit is disposed as a polarizer between the optical processing module 2 and the optical detection module 4.

[0055] The polarization directions of the first processing unit and the second processing unit are perpendicular. The first processing unit and the second processing unit are one of a polarizer, a polarizing prism, or a waveplate.

[0056] Specifically, the optical computing module 3 includes one or more optical modulation units; the optical modulation unit adopts one of spatial light modulator, mask, and metasurface.

[0057] The optical modulation unit of the optical computing module 3 receives the light field processed by the optical processing module 2. By modulating the amplitude and phase information of the incident light field, it performs multiplication and addition calculations on the incident light field, simulating the function of neurons in a traditional neural network, realizing convolutional neural network calculations, and outputting the calculated light field.

[0058] In this embodiment, the spatial light modulator is a liquid crystal spatial light modulator with a pixel size of 36 μm and a resolution of not less than 380×380. It operates in amplitude modulation mode and achieves precise control of the incident light field by loading pre-trained parameters for each pixel unit.

[0059] The optical detection module 4 is a charge-coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) sensor, used to receive the output light field of the optical computing module 3, convert the optical signal into an electrical signal, use a pre-trained neural network to make classification decisions, and finally output the recognition result.

[0060] Combination Figure 3 In this embodiment, the optical detection module 4 has four detection areas on its CMOS sensor, each corresponding to one of the four types of digital recognition results. The areas are evenly distributed at the center of the sensor, and there is an appropriate interval between the areas to avoid signal interference.

[0061] In addition, the system also includes a housing module; the front end of the housing module is provided with a structure a for placing the optical input module 1; the housing module is provided with a base for fixing the optical modulation unit and the optical detection module 4; a sliding groove is provided between the base of the optical modulation unit and the optical detection module 4 for adjusting the distance between the optical modulation unit and the optical detection module 4.

[0062] Combination Figure 4 Part a is a rectangular area used to place the OLED screen. Part b is the base of the spatial light modulator (SLM), with screw holes at the bottom for fixing the SLM. Part c shows the SLM base and the slide groove for fixing the camera mold, which allows the camera and SLM to slide in the x-direction. Part d is the camera fixing mold, with a circular opening at the front for mounting the telecentric lens; multiple screw holes on the back for front-rear fixing of the camera using rubber pads; and countersunk holes on the sides for left-right fixing of the camera.

[0063] Combination Figure 6 , Figure 7 and Figure 8 The main body of the mold consists of four long rectangular plates forming the optical transmission and computing channel, including side panels, a bottom plate, and a top plate. The side panels have multiple sets of opposing screw holes for fixing the SLM (Spatial Light Modulator). Rubber pads are attached to the sides of the SLM, and screws are used to secure it to the rubber pads. The side panels also feature a sliding groove structure for adjusting and fixing the camera position. The bottom plate has a set of sliding grooves and screw holes on the left side for fixing the SLM base; a set of sliding grooves on the right side is used to fix the camera base, enabling precise alignment and position adjustment. A rectangular area is cut out in the center of the top plate, with multiple sets of screw holes for leading out and fixing various power cables. The system is equipped with a power cable box with multiple sets of symmetrical hollow structures for connecting the SLM's signal and power cables, ensuring system power supply.

[0064] Combination Figure 9 The mold is equipped with two sets of SLM bases to support experiments with different network structures. The first set of bases has three slots of the same size with a certain distance between them. The protruding parts on both sides of the slots are used to fix the SLM in the front and back direction, which is suitable for three-layer network structures. The second set of bases has two slots of the same size with a larger distance between them. The SLM is also fixed by the protruding parts on both sides of the slots, which is suitable for two-layer network structures. Both sets of bases are equipped with multiple sets of screw holes for fixed connection with the base plate.

[0065] Combination Figure 10 The rear end of the mold is provided with a camera structure for fixing the light detection module 4, with appropriate space left inside to place the CMOS camera; the front end has a circular opening for installing a telecentric lens; the back end has multiple screw holes for front-to-back fixing of the camera with rubber gaskets; each side has countersunk holes for left-to-right fixing of the camera; the side of the camera box has screw holes that cooperate with the sliding grooves of the side panel to achieve precise adjustment of the camera position and ensure imaging quality.

[0066] In this embodiment, the light input module 1 uses an OLED screen as the light source; the right side of the screen displays a pre-arrayed image of the MNIST handwritten digit dataset, which is composed of multiple array sub-units to enhance the diffraction effect. The OLED screen is fixed in a rectangular space at the front end of the mold, and the back side is fixed by multiple screw holes and rubber gaskets, while the front area serves as the light input window.

[0067] In this embodiment, the light processing module 2 uses two orthogonally placed polarizers, which are respectively attached to the front and rear surfaces of the spatial light modulator, so that the spatial light modulator works in pure amplitude modulation mode, effectively filtering out ambient stray light and ensuring that the light in the optical path is linearly polarized, thus providing stable conditions for subsequent diffraction calculations.

[0068] In this embodiment, the optical computing module 3 employs a liquid crystal spatial light modulator (SLM). The SLM is fixed to the second set of bases (double-layer network structure bases) inside the mold, and the bases are connected to the bottom plate via screw holes. The side of the SLM is connected to the side panel via multiple sets of screw holes, and the screws abut against the rubber pads on the side of the SLM for stable fixation. The system employs a single-layer amplitude modulation diffraction neural network.

[0069] In this embodiment, the light detection module 4 uses a combination of a commercial CMOS camera and a telecentric lens. The magnification of the lens does not change with the distance between the object and the lens. The camera is fixed in a dedicated structure at the rear of the mold. The telecentric lens is precisely installed through a circular opening at the front end, and the rear screw holes are secured to the camera front-to-back with rubber gaskets. The side countersunk holes secure the camera left-to-right. The camera housing is connected to the side panel slide groove through side screw holes, allowing for precise adjustment of the imaging position to ensure that the SLM-modulated light field is accurately imaged onto the CMOS sensor.

[0070] In this embodiment, the neural network is trained using a modified MNIST dataset to perform a four-class digit classification task. The training data is preprocessed, expanded to an appropriate resolution, and consists of sub-units of multiple arrays. The network uses amplitude modulation, and the network parameters are optimized using a backpropagation algorithm. After training, the network parameters are imported into SLM for actual optical inference testing.

[0071] In this embodiment, the CMOS sensor of the light detection module 4 is provided with 4 detection areas, which correspond to 4 types of digital recognition results. Each area is evenly distributed at the center of the sensor and an appropriate interval is maintained between the areas to avoid signal interference.

[0072] An experiment was conducted based on the scheme of this embodiment, and 200 sets of test images (50 images per class) were identified. The experimental results are shown in the table below:

[0073] The prediction is 0. The prediction is 1 The prediction is 3 The prediction is 7. result Number 0 50 0 0 0 correct Number 1 0 50 0 0 correct Number 3 10 16 18 6 correct Number 7 2 2 0 46 correct

[0074] Experimental results show that 82% of the test samples were correctly identified, approaching the theoretical accuracy of 90.27% in the simulated environment. Analysis indicates that identification errors mainly stemmed from optical alignment errors and stray light interference in the experimental environment. However, the overall results fully validate the effectiveness and practicality of the portable diffractive photonic neural network system. This system, through a compact 3D-printed mold design, achieves precise fixation and alignment of optical components, providing a reliable platform for the practical application of diffractive neural networks.

[0075] The embodiments described above are merely one implementation method of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A portable diffractive neural network image recognition system, characterized in that, The system comprises a light input module, a light processing module, a light calculation module and a light detection module. The light input module is connected with the light processing module and the light calculation module, and the light calculation module is connected with the light processing module and the light detection module. The light input module is used to obtain a spectral image light field, the light processing module is used to process the obtained spectral image light field, the light calculation module is used to perform diffraction neural network calculation on the light field, and the light detection module is used to classify the calculation result, thereby outputting an image recognition result.

2. The portable diffractive neural network image recognition system of claim 1, wherein, The light input module obtains a spectral image light field with a certain light intensity distribution by using an optical element, and is composed of a plurality of array units and adopts an OLED screen or an LED array panel.

3. The portable diffractive neural network image recognition system of claim 1, wherein, The light processing module comprises a first processing unit and a second processing unit, which are used to filter ambient stray light and make the light beam in the light path linearly polarized light. The first processing unit is arranged between the light input module and the light calculation module as a polarizer, and the second processing unit is arranged between the light processing module and the light detection module as an analyzer. The polarization directions of the first processing unit and the second processing unit are perpendicular to each other.

4. The portable diffractive neural network image recognition system of claim 3, wherein, The first processing unit and the second processing unit are polarizing plates, polarizing prisms or wave plates.

5. The portable diffractive neural network image recognition system of claim 1, wherein, The light calculation module comprises one or more optical modulation units. The optical modulation unit adopts one of a spatial light modulator, a mask plate and a metasurface.

6. The optical modulation unit receives the light field processed by the light processing module, modulates the amplitude information of the incident light field, realizes the multiplication and addition calculation in the diffraction neural network, completes the diffraction propagation and interference of the light field, and outputs the calculated light field distribution.

7. The portable diffractive neural network image recognition system of claim 5, wherein, The spatial light modulator adopts a liquid crystal type spatial light modulator with a pixel size of 36μm and a resolution of not less than 380×380, and works in amplitude modulation mode. The incident light field is accurately controlled by loading pre-training parameters for each pixel unit.

8. The portable diffractive neural network image recognition system of claim 1, wherein, The light detection module is a charge coupled device (CCD) or a complementary metal oxide semiconductor (CMOS) for receiving the output light field of the light calculation module, converting the light signal into an electric signal, classifying the signal by using a pre-trained neural network, and finally outputting the recognition result.

9. The portable diffractive neural network image recognition system of claim 5, wherein, The system further comprises a housing module. The front end of the housing module is provided with a structure a for placing the light input module. A base for fixing the optical modulation unit and the light detection module is arranged in the housing module. A sliding groove is arranged between the base of the optical modulation unit and the base of the light detection module, for adjusting the distance between the optical modulation unit and the light detection module.

Images