A flying target recognition system based on photonic neural network and its construction method

Abstract

The present invention discloses a flight target recognition system based on a photonic neural network and a construction method. The ultrafast flight target recognition system based on a photonic neural network is designed and constructed for high-speed moving aircraft, including but not limited to airplanes, drones, helicopters, gliders, delta wings and other aircraft, as well as missiles, to achieve rapid recognition of high-speed moving targets. The present invention provides a flight target recognition system based on a photonic neural network, including an imaging and filtering module, a light field scaling and coupling module, a photonic neural network module, a mode control module, and an output light field coupling module; a construction method for a flight target recognition system, including artificial neural network design and photonic implementation of the artificial neural network; the present invention designs a photonic neural network transmission and control model based on all-optical passive components, trains for small sample learning of high-speed flying targets, builds static and dynamic target recognition systems, and realizes recognition of high-speed flying moving targets.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, in particular to machine vision and target recognition technology, and more specifically to a flight target recognition system and construction method based on a photonic neural network, which can be used in industrial production, modern logistics, disaster warning, and national security. Background Art

[0002] Target recognition technology utilizes photoelectric detection equipment and computers to identify distant targets. The computer extracts target characteristics from the signals acquired by the photoelectric detection equipment and calculates target situation estimation parameters. Finally, a classifier uses a discriminant function determined by a large number of training samples to make identification decisions. Target recognition technology plays an important role in industrial production, modern logistics, disaster warning, port security, and military fields.

[0003] The core of traditional machine vision systems is photoelectric sensors and computers. Their processing speed is limited by the sensor detection rate, data transmission speed, and algorithm computation speed. Due to the constraints of the physical architecture, significant breakthroughs and improvements in processing speed are difficult to achieve. Object recognition technology based on traditional architectures suffers from slow recognition speed and poor timeliness in applications. This shortcoming is particularly pronounced in high-intensity combat environments, such as the military. Furthermore, object recognition algorithms require high-performance computers for processing, which are generally large and power-hungry. This significantly limits their application, especially in military applications, where they cannot be installed on unmanned aerial vehicles, which require a small size, low weight, and low power consumption. The proposed photonic neural network can effectively address this problem.

[0004] Unlike most current computer-based artificial neural networks, photonic neural networks (PNs) are a type of artificial neural network (ANN) based on optical components. They transcend the physical architecture of target recognition systems by directly acquiring target classification and situational awareness information through photon operations, completing target recognition at the speed of light. Their compact size, light weight, and low power consumption make them a promising alternative to traditional computer-based target recognition systems, enabling ultrafast target recognition. Current PN target recognition systems based on optical components fall into three categories: photonic chips, diffraction-based PNs, and scattering-medium PNs. Photonic chips offer high integration but cannot directly process optical images, requiring optical-to-electrical-to-optical conversion, resulting in low efficiency and significant fabrication complexity. Diffraction-based PNs are simple and easy to implement, but lack flexibility, and their system size increases with network complexity. Scattering-medium PNs, on the other hand, offer compact structures but are challenging to design and fabricate. Therefore, a PN target recognition system that combines these advantages to address the slow recognition speed, bulk, and high power consumption of traditional target recognition systems remains lacking. Summary of the Invention

[0005] The present invention aims to provide a flying target recognition system based on a photonic neural network. It is designed and constructed for high-speed moving aircraft, including but not limited to airplanes, drones, helicopters, gliders, delta wings and other aircraft, as well as military aircraft with missiles and guided weapons, to achieve rapid recognition of high-speed moving targets.

[0006] The present invention provides a flying target recognition system based on a photonic neural network, comprising an imaging and filtering module, a light field scaling and coupling module, a photonic neural network module, a mode control module, and an output light field coupling module;

[0007] The imaging and filtering module performs imaging and filtering on the target;

[0008] The light field scaling and coupling module is a 4F system including an objective lens, the optical axis of the imaging and filtering module and the optical axis of the light field scaling and coupling module need to be coaxial, and the image plane of the imaging and filtering module and the end face of the photonic neural network module are conjugate surfaces to each other;

[0009] The photonic neural network module is used to operate the light field input by the light field scaling and coupling module, and includes a linear operation part and a nonlinear operation part. The linear operation part uses a multimode optical waveguide to form a photonic neural network;

[0010] The output light field coupling module is used to receive the light field after operation by the photonic neural network module and output parallel operation results for the light field.

[0011] As a further improvement of the present invention: the filter in the imaging and filtering module adopts a tunable filter or a narrow-band filter, the central wavelength of which is any wavelength from visible light to mid-infrared and the bandwidth should be less than 2nm.

[0012] As a further improvement of the present invention: the photonic neural network adopts a multimode optical waveguide, including a multimode optical fiber, a multi-core optical fiber or an on-chip multimode waveguide, and the photonic neural network has a reconfigurable characteristic.

[0013] As a further improvement of the present invention: the linear operation of the photonic neural network is realized through multiple mode control modules.

[0014] As a further improvement of the present invention: the nonlinear operation of the photonic neural network is achieved through mode nonlinear coupling, including four-wave mixing, cross-phase modulation or phonon-photon coupling.

[0015] As a further improvement of the present invention: the output light field coupling module includes light field coupling and photoelectric conversion, the light field coupling includes free space coupling, an imaging system or a photon lantern; the photoelectric conversion includes a planar array detector, a line detector array or multiple point detectors.

[0016] The present invention also provides a method for constructing a flying target recognition system, including an artificial neural network design and a photonics implementation of the artificial neural network;

[0017] The artificial neural network design includes:

[0018] Using different small sample learning methods, we compared the complexity of each network structure and the degree of demand for high-speed parallel computing to obtain the most suitable artificial neural network structure for optical implementation.

[0019] Acquire a data set and preprocess the samples; input the preprocessed samples into the artificial neural network for learning, obtain test results, adjust parameters to optimize the structure of the artificial neural network, and verify through a computer to obtain the network structure and network parameters of the artificial neural network;

[0020] The photonics implementation of the artificial neural network includes:

[0021] The artificial neural network is implemented using all-optical passive components; the mode coupling matrix is ​​calculated using the optimized artificial neural network structural parameters to determine the parameters of the mode control module; the flight target recognition system based on the photonic neural network is constructed; and the flight target recognition system based on the photonic neural network is experimentally verified and parameter optimization and adjustment are performed.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] The present invention provides a flight target recognition system based on a photonic neural network. It designs a photonic neural network transmission and control model based on all-optical passive components, conducts training for small sample learning of high-speed flight targets, builds static and dynamic target recognition systems, and realizes the recognition of moving targets in the high-speed flight field. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A flow chart of a method for building a flying target recognition system based on a photonic neural network according to an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of the structure of a flying target recognition system based on a photonic neural network provided by an embodiment of the present invention;

[0026] In the figure: imaging and filtering module 1, light field scaling and coupling module 2, photonic neural network module 3, mode control module 4, output light field coupling module 5. DETAILED DESCRIPTION

[0027] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0028] While currently widely used deep learning methods rely heavily on massive amounts of labeled data, this is often difficult to obtain in many fields, such as industrial production, military security, and high-tech research and development. Few-shot learning, with its ability to learn new categories using minimal labeled samples without changing the training model, has become a key approach to addressing these challenges.

[0029] The embodiment of the present invention physically realizes an all-optical component using an artificial neural network structure based on small sample learning, combines small sample learning with a photonic neural network, and provides an ultra-fast recognition system for unknown high-speed flying moving targets.

[0030] See also Figure 1 The flowchart of the method for building a flying target recognition system based on a photonic neural network of the present invention is shown, which includes the artificial neural network design S1 and the photonic implementation of the artificial neural network S2. The specific process is as follows:

[0031] The artificial neural network design S1 specifically includes the following steps:

[0032] Step S101: Research and compare the test results of different small sample learning methods, such as twin networks, matching networks, prototype networks, graph neural networks, and meta-learning optimizers, on the miniImageNet dataset. The results show that the meta-learning optimizer method has high computational cost and slow training speed in electronic neural networks, and has low structural complexity, making it very suitable for implementation in high-speed parallel photonic neural networks. This embodiment uses a meta-learning optimizer as the artificial neural network structure for aircraft target feature extraction and recognition. Combined with the photonic implementation mechanism, its feasibility is explored, and the network structure is adjusted. Ultimately, a photonic neural network structure suitable for aircraft target recognition that can be implemented using photonic components is determined.

[0033] Step S102: Select an aircraft sample data set. The data set includes but is not limited to aircraft such as airplanes, drones, helicopters, gliders, delta wings, and military aircraft such as missiles and guided weapons. This embodiment uses the Caltech aircraft side image data set from California Institute of Technology. The samples are preprocessed, including de-meaning, normalization, decorrelation, and whitening. The de-meaning is used to center each dimension of the input data to zero, the normalization is used to normalize the amplitude to the same range to reduce the interference caused by the difference in the value range of each dimensional data, the decorrelation is used for dimensionality reduction, and the whitening is used to normalize the amplitude on each feature axis of the data.

[0034] Step S103: Select 20 preprocessed samples and divide them into 80% training set and 20% test set, use the training set to input the meta-learning artificial neural network for learning, test the results through the test set, adjust the network parameters to optimize the network structure, and verify its reliability through computer, including adversarial samples, robustness, interpretability, etc.

[0035] Through computer-aided design of artificial neural networks, the network structure and network parameters of the artificial neural network suitable for aircraft targets are obtained, and are implemented through photonic components, that is, the photonic implementation S2 of the artificial neural network.

[0036] The photonics implementation S2 of the artificial neural network specifically includes the following steps:

[0037] Step S201: Implement the artificial neural network using all-optical passive components; photonics implement the artificial neural network structure designed in artificial neural network design S1. The linear operation portion of the traditional electronic neural network is replaced by a multimode optical waveguide. To achieve high-speed parallel optical linear operations, the photonic neural network operation core uses a multimode optical waveguide, including but not limited to large-diameter multimode optical fiber, multi-core optical fiber, integrated on-chip multimode waveguide, etc. In this embodiment, a large-diameter 105μm multimode optical fiber is used. The nonlinear operation portion is mainly implemented by nonlinear coupling between modes, including but not limited to four-wave mixing, cross-phase modulation, phonon-photon coupling, active absorption, and frequency up-conversion. In this embodiment, cross-phase modulation is used. Parameter control such as feedback and weight bias uses multimode optical waveguides and long-period gratings to achieve control in both wavelength and spatial dimensions. Combined with a directional bending system, it is applied to the transmission system to achieve reconfigurable inter-mode control and achieve feedback iterative adjustment of weights and biases.

[0038] Step S202: Calculate the mode coupling matrix in the multimode optical fiber using the artificial neural network structure parameters in the artificial neural network design S1, adjust the detection mode coupling matrix, and determine the parameters of each unit in the mode control module.

[0039] Step S203: Build a flying target recognition system based on photonic neural network, such as Figure 2 The figure shows the structure diagram of the flight target recognition system based on photonic neural network, which includes the following modules: imaging and filtering module 1, light field scaling and coupling module 2, photonic neural network module 3, mode control module 4, and output light field coupling module 5. Among them, the imaging and filtering module 1 images and filters the aircraft target, trying to extract the wavelength frequency characteristics of the object to be measured. The filter in the imaging and filtering module 1 adopts an adjustable filter or a narrowband filter, the center wavelength of which can be any wavelength from visible light to mid-infrared and the bandwidth should be less than 2nm. This embodiment uses an adjustable filter; the light field scaling and coupling module 2 is a 4F system including an objective lens. The optical axes of the two must be coaxial, so that the image plane of the imaging and filtering module 1 and the end face of the photonic neural network module 3 are conjugate surfaces to each other; the photonic neural network module 3 is the linear and nonlinear operation part optically implemented in the above steps, which is used to operate on the input light field; the mode control module 4 is the mode control part in the above steps, which is used to feedback and iteratively adjust the photonic neural network parameters; the output light field coupling module 5, after the light field is transmitted by the photonic neural network module 3, enters the output light field coupling module 5, thereby outputting parallel operation results, including light field coupling and photoelectric conversion. The light field coupling methods include but are not limited to free space coupling, imaging system, and photon lantern. This embodiment uses an imaging system. Photoelectric conversion includes but is not limited to a planar array detector, a line detector array, and a plurality of point detectors. This embodiment uses a planar array detector.

[0040] Step S204: Using the photonic neural network-based flight target recognition system to identify static aircraft and high-speed moving aircraft targets, adjusting parameters to optimize the photonic neural network structure, and finally completing the optimization of the target recognition system based on the multimode optical waveguide photonic neural network.

[0041] It should be noted that, in this document, relational terms such as first and second, etc., are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "include," "comprise," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. Without further limitation, elements defined by the phrase "include..." or "comprising..." do not exclude the presence of additional elements in the process, method, article, or terminal device comprising the elements. Furthermore, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the number itself; "above," "below," "within," etc., are understood to include the number itself.

[0042] Although the above embodiments have been described, those skilled in the art may make additional changes and modifications to these embodiments once they know the basic creative concepts. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present invention.

[0043] The above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A flying target recognition system based on a photonic neural network, characterized by: Including imaging and filtering module, light field scaling and coupling module, photonic neural network module, mode control module, and output light field coupling module; The imaging and filtering module performs imaging and filtering on the target; The light field scaling and coupling module is a 4F system including an objective lens, the optical axis of the imaging and filtering module and the optical axis of the light field scaling and coupling module need to be coaxial, and the image plane of the imaging and filtering module and the end face of the photonic neural network module are conjugate surfaces to each other; The photonic neural network module is used to operate the light field input by the light field scaling and coupling module, and includes a linear operation part and a nonlinear operation part. The linear operation part uses a multimode optical waveguide to form a photonic neural network; The nonlinear operation of the photonic neural network is achieved through mode nonlinear coupling, including four-wave mixing, cross-phase modulation or phonon-photon coupling; The output light field coupling module is used to receive the light field after the operation of the photonic neural network module and output the parallel operation results of the light field; The output light field coupling module includes light field coupling and photoelectric conversion. The light field coupling includes free space coupling, imaging system or photon lantern; the photoelectric conversion includes a planar array detector, a line detector array or multiple point detectors.

2. The flying target recognition system based on a photonic neural network according to claim 1, characterized in that: The filter in the imaging and filtering module is a tunable filter or a narrow-band filter, the central wavelength of which is any wavelength from visible light to mid-infrared and the bandwidth should be less than 2nm.

3. The flying target recognition system based on photonic neural network according to claim 1, characterized in that: The photonic neural network adopts a multimode optical waveguide, including a multimode optical fiber, a multi-core optical fiber or an on-chip multimode waveguide, and the photonic neural network has a reconfigurable characteristic.

4. The flying target recognition system based on photonic neural network according to claim 1, characterized in that: The linear operation of the photonic neural network is achieved through multiple mode control modules.

5. A method for constructing a flying target recognition system, the method being used for constructing a flying target recognition system based on a photonic neural network as claimed in any one of claims 1 to 4, characterized in that: Including artificial neural network design and photonic implementation of artificial neural networks; The artificial neural network design includes: Using different small sample learning methods, we compared the complexity of each network structure and the degree of demand for high-speed parallel computing to obtain the most suitable artificial neural network structure for optical implementation. Acquire a data set and preprocess the samples; input the preprocessed samples into an artificial neural network for learning, obtain test results, adjust parameters to optimize the structure of the artificial neural network, and verify through a computer to obtain the network structure and network parameters of the artificial neural network; The photonics implementation of the artificial neural network includes: The artificial neural network is implemented using all-optical passive components; the mode coupling matrix is ​​calculated using the optimized artificial neural network structural parameters to determine the parameters of the mode control module; the flight target recognition system based on the photonic neural network is constructed; and the flight target recognition system based on the photonic neural network is experimentally verified and parameter optimization and adjustment are performed.

Images