A neural network-based method and device for detecting extremely low-power optical targets

Abstract

The present invention relates to the field of target detection, and in particular to an extremely low-power optical target detection method based on a neural network. The method comprises the following steps: constructing a convolutional neural network to extract target image features; normalizing the extracted target image features to obtain target samples; finding convolution kernel weights, selecting a convolutional neural network framework with the highest average recognition accuracy after training, building a learnable optical prism device and applying it to a standard target detection database, learning to obtain the convolution kernel weights of each layer, finding the threshold of the excitation layer of the convolutional neural network framework, and selecting the excitation function with the highest target position accuracy after training and testing; determining the dispersion coefficient of the optical prism and the threshold of the filter film, customizing a scattering mirror, forming a target light spot through the scattering mirror, and finally imaging through a camera. By using several layers of optical prisms of the convolutional neural network to calibrate the target frame, the method achieves high real-time efficiency and accuracy without consuming any battery energy.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to an extremely low-power optical target detection method and device based on a neural network. Background Art

[0002] Target detection is a relatively popular technology. This method obtains specified image information in a specific way and marks the position of the located target on the image. The target detection system is centered on target recognition technology and is a high-tech technology that is being tackled in the current international scientific and technological field. The great breakthroughs made by neural networks in the field of image recognition have promoted the rapid progress of target detection as an image application, provided higher stability for target detection, and thus promoted the widespread application of target detection in more and more fields such as entertainment and security. At the same time, it also puts higher requirements on target extraction. Target detection needs to have higher real-time and high efficiency, and be able to minimize power consumption. To this end, the present invention proposes an extremely low-power optical target detection method and device based on neural networks. Summary of the Invention

[0003] The present invention provides an extremely low-power optical target detection method and device based on a neural network, aiming to solve the problem in the prior art of lacking a method that can comprehensively, automatically, accurately, and extremely low-power extract target image features and perform identification.

[0004] The technical solution adopted by the present invention is:

[0005] A neural network-based method for detecting optical targets with extremely low power consumption comprises the following steps:

[0006] a. Construct a convolutional neural network and feed each image in the target database into the constructed convolutional neural network to extract target image features;

[0007] b. Normalize the extracted target image features and perform affine projection into a low-dimensional space to obtain a projection matrix. Train the projection matrix using a loss function to obtain target samples for each image.

[0008] c. Find the convolution kernel weight value and select the convolutional neural network framework with the highest average recognition accuracy after training. The structure of the convolutional neural network framework includes: light input layer-convolution layer-excitation function-convolution layer-excitation function-convolution layer-excitation function-fully connected layer-light output layer;

[0009] d. Build a learnable optical prism device and apply it to a standard target detection database. The features of the target sample to be detected are converted into light with target information through the optical prism device and then injected into the convolutional neural network framework for training. The weights of each convolution kernel layer are learned. The threshold of the excitation layer of the convolutional neural network framework is determined. After training and testing, the excitation function with the highest target location accuracy is selected.

[0010] e. Set the dispersion coefficient of the optical prism and the threshold of the filter film in the optical prism framework according to the convolution kernel weights and excitation functions of each layer. Customize the scattering mirror according to the dispersion coefficient of each layer of the optical prism. Dispersion and filtering are performed through the optical prism and filter film respectively. The target light spot is formed by the scattering mirror, and finally the image is formed through the camera.

[0011] After obtaining the convolution kernel weights of each layer in step d, verify the target detection accuracy. If it is greater than or equal to the set threshold, save the model parameters trained in step d; if it is less than the set threshold, re-execute step c.

[0012] The accurate convolution kernel weights are obtained by combining forward propagation and back propagation:

[0013] Forward propagation: The target image enters the convolutional neural network framework from the light input layer, performs a weighted sum operation with the weight of the corresponding convolution kernel, and the bias term value is 0. It then passes through an excitation layer, and the final result is the output result of this layer; if the actual output of the output layer is the same as the expected output, the learning ends and the convolution kernel weight is saved; if the output result obtained by the output layer is different from the expected output, the error back propagation process is switched;

[0014] Back propagation: The difference between the actual output and the expected value is calculated by backpropagating the original convolution layer channel, and then backpropagated to the light input layer through the convolution layer. During the backpropagation process, the error is distributed to each unit in each layer to obtain the error signal of each single layer, and it is used as the basis for correcting the weights of each unit; after continuously adjusting the convolution kernel weights and thresholds of each layer, the error signal is reduced to a minimum.

[0015] When mapped into high-dimensional input during the back-propagation process, the convolution kernel is flipped 180 degrees according to the deconvolution principle, and the target information is gradually extracted. The optical prism changes the prism angle according to the convolution kernel weight after flipping 180 degrees to obtain the dispersion coefficient of each layer of prism.

[0016] After the light with target information enters the optical prism, it undergoes light dispersion operations with different weights at each layer, and each layer obtains the corresponding target information light according to different dispersion coefficients; the filter film screens the light scattered by the optical prism and filters out the light outside the threshold; the scattering mirror forms a target light spot at the target position based on the dispersed light containing target information.

[0017] After the target light spot is formed, it is mapped back to the original image according to the dispersion coefficient through a special scattering mirror, and the target position is calibrated to form a target frame.

[0018] The target database may be a CASIA-WebFace face database or an original face image acquired through a development interface of an electronic terminal and collected by an image acquisition unit of an electronic terminal device.

[0019] A neural network-based, ultra-low-power optical target detection device includes an optical prism device that can be installed at the front lens position of an electronic device. The optical prism device includes multiple layers of optical prisms, multiple layers of filter films, and a scattering mirror with a specific scattering rate. The prisms are stacked layer by layer. The number of filter films is determined according to an excitation function and is embedded behind the optical prism. The scattering mirror is embedded behind the last layer of optical prism.

[0020] The optical prism is made of rigid optical glass or alkali metal halide crystal material with variable dispersion coefficient, the filter film is made of glass crystal or silicon carbide material using ion amplitude dielectric film process technology, and the scattering mirror is made of coated silicon wafer.

[0021] The optical prism device is followed by a beam splitter.

[0022] Compared with the existing technology, the beneficial effects of the present invention are: target detection is performed through operations such as light dispersion and filtering, the target light spot is directly used as the output, and autonomous learning is performed through training sample data. At the same time, it has the feature of sharing the weights of each pixel in each layer, reducing the complexity of the framework. The operation of the filter film enhances the robustness of the light spot, so that it can better receive the target light under different light conditions. The present invention does not consume any power other than light information, and at the same time needs to improve the sensitivity of camera imaging. The present invention uses a multi-layer optical prism structure similar to a convolutional neural network to calibrate the target frame. The prism structure is not large, and it is easy to apply to small data sets. It has high real-time and accuracy, and does not consume any battery energy. It can greatly improve the shooting time and service life of cameras and other electronic devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A model structure diagram of an optical prism device according to an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of a convolutional neural network according to an embodiment of the present invention;

[0025] Figure 3 2 is a block diagram of an optical face detection system based on a convolution process in an embodiment of the present invention;

[0026] Figure 4This is a flow chart of a method for detecting a face frame using an optical low-power camera in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The following is a further description of specific embodiments of the present invention in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is intended to facilitate understanding of the present invention and does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0028] Reference Figure 1-4 , an extremely low power consumption optical target detection method based on a neural network, comprising the following steps:

[0029] a. Construct a convolutional neural network and feed each image in the target database into the constructed convolutional neural network to extract target image features;

[0030] b. Normalize the extracted target image features and perform affine projection into a low-dimensional space to obtain a projection matrix. Train the projection matrix using a loss function to obtain target samples for each image.

[0031] c. Find the convolution kernel weight value and select the convolutional neural network framework with the highest average recognition accuracy after training. The structure of the convolutional neural network framework includes: light input layer-convolution layer-excitation function-convolution layer-excitation function-convolution layer-excitation function-fully connected layer-light output layer;

[0032] d. Build a learnable optical prism device and apply it to a standard target detection database. The features of the target sample to be detected are converted into light with target information through the optical prism device and then injected into the convolutional neural network framework for training. The weights of each convolution kernel layer are learned. The threshold of the excitation layer of the convolutional neural network framework is determined. After training and testing, the excitation function with the highest target location accuracy is selected.

[0033] e. The dispersion coefficient of the optical prism and the threshold of the filter film in the optical prism framework are set according to the convolution kernel weights and excitation functions of each layer. A scattering mirror is customized according to the dispersion coefficient of each layer of the optical prism. Dispersion and filtering are performed respectively through the optical prism and the filter film, and the target light spot is formed through the scattering mirror. Thus, the target light spot is formed before the light reaches the camera screen, and finally the image is formed through the camera to ensure the efficiency and real-time performance of the camera.

[0034] After obtaining the convolution kernel weights of each layer in step d, verify the target detection accuracy. If it is greater than or equal to the set threshold, save the model parameters trained in step d; if it is less than the set threshold, re-execute step c.

[0035] Convolutional neural networks are a type of feedforward neural network that includes convolution calculations and has a deep structure. They have representational learning capabilities, can translate and classify input information by changing weights according to a hierarchical structure, and can determine the output of neurons through the threshold of the excitation function. Studies have shown that the accuracy of target detection can be extracted through trained convolutional neural networks. Studies have found that during the propagation of light, prisms have a similar effect to convolution layers. The target light spot can be extracted through optical prism dispersion, filtering and other operations without consuming any energy other than light. Based on this, the present invention performs target detection through operations such as light dispersion and filtering, directly uses the target light spot as the output, and performs autonomous learning through training sample data. At the same time, it has the feature of sharing the weights of each pixel in each layer, reducing the complexity of the framework. The operation of the filter film enhances the robustness of the light spot, so that it can better receive the target light under different light conditions. The present invention does not consume any power except for light information, and at the same time needs to improve the sensitivity of camera imaging. The present invention uses a multi-layer optical prism structure similar to a convolutional neural network to calibrate the target frame. The prism structure is small, easy to apply on small data sets, has high real-time performance and accuracy, and does not consume any battery energy. It can greatly improve the shooting time and service life of cameras and other electronic devices.

[0036] The target database includes a large number of original target images, which can be directly obtained from the CA SIA-WebFace face database or original face pictures obtained through the development interface of the electronic terminal and collected by the image collection unit of the electronic terminal device.

[0037] Taking the CASIA-WebFace face database for face target detection as an example, the process includes the following steps:

[0038] Step 1: Divide the facial images in the CASIA-WebFace face database into three categories: training samples, test samples, and validation samples. Feed the training set in the face database into the constructed convolutional neural network to extract features and read the training data.

[0039] Step 2: Normalize the face image after reading the training data in step 1;

[0040] Normalization is performed on the incoming light from the facial image. This is because neural networks train and predict based on the statistical distribution of samples in an event. The face's location is determined by statistical analysis and training on the training set images. Normalization is a statistical probability distribution between 0 and 1. When all sample input signals are positive, the weights corresponding to the convolution kernel can only increase or decrease simultaneously, which results in slow learning. To avoid this and speed up neural network learning, the incoming signal can be normalized so that the mean of all sample input signals is close to 0 or smaller than its mean square error. This is why the normalization step is added.

[0041] Step 3: Construct the convolutional neural network framework structure, which is: light input layer-convolution layer-excitation function-convolution layer-excitation function-convolution layer-excitation function-full connection layer-light output layer;

[0042] Step 4: Place the training sample after the normalization processing in step 2 into the framework structure constructed in step 3 for training. The training process is forward propagation and back propagation. Place the verification sample after the normalization processing in step 2 into the framework structure constructed in step 3 for verification. If the accuracy of the test result of the verification sample is greater than or equal to the set threshold, execute step 5. If the test result of the verification sample is less than the set threshold, re-execute step 3 (i.e., reselect the convolutional neural network framework structure).

[0043] Step 5: Save the model parameters trained in step 4, and use the trained model parameters to test the test samples normalized in step 2 to obtain the model detection results;

[0044] Step 6: Determine the prism frame structure according to the convolutional neural network framework structure and the convolution kernel weight. The structure is: light input layer-optical prism layer-optical prism layer-optical prism layer-scattering mirror layer-light output layer; the optical prism layer includes (optical prism-filter film).

[0045] The following further describes the above embodiment by combining specific calculation formulas and taking face detection based on a convolutional neural network as an example:

[0046] (1) Selecting a facial image database and preprocessing facial images

[0047] Li Ziqing's CASIA-WebFace face database was chosen. It contains over 500,000 facial images, including 10,575 faces. The number of facial images per category ranges from dozens to hundreds, making it a highly suitable database for training face detection networks. Image preprocessing begins by detecting facial landmarks in the database's original images of varying sizes, classifying and detecting the facial image locations. The convolution kernel weights are then adjusted based on the facial image locations.

[0048] (2) Construction of optical prism device

[0049] The optical prism device consists of at least nine optical prism layers, each followed by a filter film. After light enters the input layer, it passes through the optical prism layer to extract and verify facial information, and finally passes through the scattering mirror to the output layer. The following is a detailed introduction to the construction principle of the prism device:

[0050] Optical prism layer operation: The formula is as follows:

[0051]

[0052] in, represents the jth beam point of the tth layer, M j represents the selected set of input rays, represents the offset value corresponding to the j-th beam of light (in the present invention, the offset value is 0), Represents the weight between the jth beam of light in the tth layer and the i-th prism pixel in the t+1th layer. "*" indicates a convolution-like operation. The learnable optical prism device is arranged as light input layer-optical prism layer-optical prism layer-optical prism layer-scattering mirror layer-light output layer, which is similar to the unsaturated gradient and fast calculation speed characteristics of the ReLu function. In the selection of the filter film, a characteristic similar to the activation function ReLu is selected, such as Figure 2 As shown, the current output is represented as:

[0053] x e =f(u e )

[0054] u e =W e x e-1 +b e

[0055] where x e Represents the output of the current layer, u e represents the input of the filter (the result of the weight calculation of the current layer), f() represents the filter activation function, W e is the weight of the current layer, b e A bias can be added (can be selected as 0 in the filter film).

[0056] Other steps and parameters are the same as those in the above embodiment.

[0057] Taking the example of face target detection using the original face image database, the process includes the following steps:

[0058] (1) Each image in the face database is fed into the three constructed convolutional neural networks to extract features;

[0059] (2) Normalize the output features and affine project them into a low-dimensional space to obtain a projection matrix. The projection matrix is ​​trained through a loss function to obtain the face position of each image.

[0060] (3) Find the convolution kernel weight value and select the convolutional neural network framework with the highest average recognition accuracy after training;

[0061] (4) Apply the selected learnable optical prism framework to the standard face detection database, project the characteristic light of the face image to be detected into the trained framework, find the threshold of the excitation layer, and select the excitation function with the highest face position accuracy after training and testing;

[0062] (5) According to the weights of the convolution kernels and the excitation function of each layer, the prism dispersion coefficient and the filter threshold are set; according to the dispersion coefficient of the prism device of each layer, the scattering mirror is customized to form the face light spot.

[0063] Among them, the three convolutional neural network frameworks constructed in step (1) specifically include: convolutional neural network framework A, convolutional neural network framework B, and convolutional neural network framework C; convolutional neural network framework A includes 9 modules (i.e., 9 prism layers): each module first passes through an optical prism layer and then applies a corrective filter film; convolutional neural network framework B has one more prism layer than the convolutional neural network framework A; convolutional neural network framework C has two more prism layers than convolutional neural network framework A.

[0064] The prediction and recognition accuracy of the three convolutional neural network frameworks are shown in the following table:

[0065]

[0066] In step (3), after training and testing, the convolutional neural network framework with the highest average recognition accuracy is selected. Specifically, the training sets of the three convolutional neural network frameworks are removed, and the recognition accuracy of the three frameworks is tested through the test set.

[0067] In order to obtain accurate convolution kernel weights, a combination of forward propagation and back propagation is used. Forward propagation: the target image enters the convolutional neural network framework from the light input layer, and is weighted and calculated with the corresponding convolution kernel weights, with the bias term value being 0. Then it passes through a layer of excitation layer, and the final result is the output result of this layer; if the actual output of the output layer is the same as the expected output, the learning ends and the convolution kernel weights are saved; if the output result obtained by the output layer is different from the expected output, the error back propagation process is switched to. Back propagation: the difference between the actual output and the expected value is calculated by back propagating the original convolution layer channel, and back propagated through the convolution layer to the light input layer. During the back propagation process, the error is distributed to each unit of each layer, and the error signal of each single layer is obtained, which is used as the basis for correcting the weights of each unit; after continuously adjusting the convolution kernel weights and thresholds of each layer, the error signal is reduced to a minimum.

[0068] The process of continuous adjustment of weights and thresholds is the learning and training process of convolutional neural networks. After forward propagation and back propagation, the adjustment of weights and thresholds is repeated until the pre-set number of learning and training times is reached, so that the output error is reduced to an allowable range.

[0069] When mapped into high-dimensional input during the back-propagation process, the convolution kernel is flipped 180 degrees according to the deconvolution principle, and the target information is gradually extracted. The optical prism changes the prism angle according to the convolution kernel weight after flipping 180 degrees to obtain the dispersion coefficient of each layer of prism.

[0070] After entering the optical prism, light carrying target information undergoes a light dispersion operation with different weights at each layer. Each layer captures the corresponding target information light based on a different dispersion coefficient. The filter film filters the light scattered by the optical prism, filtering out light outside the threshold. The scattering mirror forms a target light spot at the target location based on the dispersed target information light. After the target light spot is formed, it passes through a special scattering mirror and maps it back to the original image based on the dispersion coefficient. The target location is then calibrated to form a target frame. For example, the captured face light spot is transformed by the scattering mirror to form a face frame, replacing the original face frame capture method, thereby reducing camera battery power.

[0071] An extremely low-power optical target detection device based on a neural network includes an optical prism device that can be installed at the front lens position of an electronic device. The optical prism device includes multiple layers of optical prisms, multiple layers of filter films, and a scattering mirror with a specific scattering rate. The prisms are stacked layer by layer. The number of filter films is determined according to the excitation function and is embedded behind the optical prism. The scattering mirror is embedded behind the last layer of optical prism. The number of prism layers and the number of small blocks in each layer are set according to the training situation. After passing through the prism layer, the light is dispersed into light with different weights (shown as lines of varying thickness in the figure). The target light is dispersed layer by layer and stacked layer by layer, and the target light information is finally extracted. The target light spot is formed by passing through the scattering mirror through the set scattering rate. To simplify the prism setting, the size of each prism layer can be consistent.

[0072] In the field of optics, prisms have effects such as dispersion and refraction on light. The filter film filters the light by setting a threshold, which is similar to the forward propagation process of a convolutional neural network. The light containing the target information passes through a multi-layer prism device after being transmitted into the photographic device, and the target information light is obtained through the specific dispersion coefficient of the prism, and then passes through the scattering mirror to form a target light spot. The prism dispersion rate is determined by the convolution kernel weights of each layer of the deconvolution neural network. The filter film is similar to the excitation function of the convolutional neural network. The scattering mirror maps the extracted target light spot back to the original image and calibrates the position of the target image in the original image.

[0073] Based on the above embodiments, the optical prism is made of rigid optical glass or alkali metal halide crystal material with changeable dispersion coefficient. The dispersion coefficient of optical glass or alkali metal halide crystal material is easy to change, highly uniform, crack-free, and has a small temperature coefficient. The filter film is made of glass crystal or silicon carbide material using ion amplitude dielectric membrane process technology. The filter film made in this way contains a specific threshold and can filter light outside the threshold. The material can also be replaced with a similar material that has the function of filtering specific light. The scattering mirror is made of a coated silicon wafer, such as an oxide silicon wafer, a low-stress SiN layer silicon wafer, a gold-plated silicon wafer, or a platinum-plated metal film silicon wafer.

[0074] Based on the above embodiment, the optical prism device is connected to a beam splitter to ensure that the target light spot does not affect the camera imaging.

[0075] This invention uses a neural network-based, ultra-low-power optical target detection method and device. Its principle is similar to the convolution process in deep learning. By constructing an optical prism device and performing large-scale training on a target dataset, the device uses the optical prism layer to perform a convolution operation, and the filter to perform an excitation function. The device extracts targets through operations such as light dispersion and filtering, which are then scattered through a scattering mirror, directly outputting the target light spot. This method improves the real-time and high-efficiency of camera target detection without consuming any power other than the light information.

[0076] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It is apparent to those skilled in the art that various changes, modifications, substitutions, and variations to these embodiments may be made without departing from the principles and spirit of the present invention, and these changes and modifications still fall within the scope of protection of the present invention.

Claims

1. An extremely low-power optical target detection method based on a neural network, characterized by: The method comprises the following steps: a. constructing a convolutional neural network, and feeding each image in the target database into the constructed convolutional neural network to extract target image features; b. Normalize the extracted target image features and perform affine projection into a low-dimensional space to obtain a projection matrix. Train the projection matrix using a loss function to obtain target samples for each image. c. Find the convolution kernel weight value, and select the convolution neural network framework with the highest average recognition accuracy after training. The structure of the convolution neural network framework includes: light input layer-convolution layer-excitation function-convolution layer-excitation function-convolution layer-excitation function-fully connected layer-light output layer; d. Build a learnable optical prism device and apply it to the standard target detection database. The features of the target sample to be detected are converted into light with target information through the optical prism device and injected into the convolution neural network framework for training to learn and obtain the weight of each layer of convolution kernel; find the threshold of the excitation layer of the convolution neural network framework, and after training and testing, select the excitation function with the highest target position accuracy; e. Set the dispersion coefficient of the optical prism and the threshold of the filter film in the optical prism framework according to the convolution kernel weight and excitation function of each layer, customize the scattering mirror according to the dispersion coefficient of each layer of optical prism, disperse and filter respectively through the optical prism and filter film, and form the target light spot through the scattering mirror, and finally form an image through the camera; Obtain accurate convolution kernel weights by combining forward propagation and backpropagation: Forward propagation: The target image enters the convolutional neural network framework from the light input layer, and is weighted and calculated with the weight of the corresponding convolution kernel, with the bias term value being 0. It then passes through an excitation layer, and the final result is the output result of this layer. If the actual output of the output layer is the same as the expected output, the learning ends and the convolution kernel weight is saved. If the output result obtained by the output layer is different from the expected output, the error back propagation process is switched to. Back propagation: The difference between the actual output and the expected value is calculated by back propagating the original convolution layer channel, and back propagated to the light input layer through the convolution layer. During the back propagation process, the error is apportioned to each unit of each layer to obtain the error signal of each single layer, which is used as the basis for correcting the weight of each unit. After continuously adjusting the convolution kernel weights and thresholds of each layer, the error signal is reduced to a minimum. When mapped to a high-dimensional input during the back propagation process, the convolution kernel is flipped 180 degrees according to the deconvolution principle, and the target information is gradually extracted. The optical prism changes the prism angle according to the convolution kernel weight after flipping 180 degrees to obtain the prism dispersion coefficient of each layer. After entering the optical prism, light carrying target information undergoes light dispersion operations with different weights at each layer, with each layer obtaining corresponding target information light according to different dispersion coefficients. The filter film screens the light scattered by the optical prism, filtering out light outside the threshold. The scattering mirror forms a target light spot at the target position based on the dispersed light containing target information. After the target light spot is formed, it is mapped back to the original image based on the dispersion coefficient through a special scattering mirror, calibrating the target position and forming a target frame.

2. The neural network-based ultra-low power optical target detection method according to claim 1, wherein: After obtaining the convolution kernel weights of each layer in step d, verify the target detection accuracy. If it is greater than or equal to the set threshold, save the model parameters trained in step d; if it is less than the set threshold, re-execute step c.

3. The neural network-based ultra-low power optical target detection method according to claim 1, wherein: The target database may be a CASIA-WebFace face database or an original face image acquired through a development interface of an electronic terminal and collected by an image acquisition unit of an electronic terminal device.

4. A device for the neural network-based ultra-low power optical target detection method according to any one of claims 1 to 3, characterized in that: The invention comprises an optical prism device that can be installed at the front lens position of an electronic device, wherein the optical prism device comprises a multi-layer optical prism, a multi-layer filter film, and a scattering mirror with a specific scattering rate; the prisms are stacked layer by layer; the number of the filter films is determined according to the excitation function and is embedded behind the optical prism; the scattering mirror is embedded behind the last optical prism layer.

5. The apparatus for the neural network-based ultra-low power optical target detection method according to claim 4, characterized in that: The optical prism is made of rigid optical glass or alkali metal halide crystal material with variable dispersion coefficient, the filter film is made of glass crystal or silicon carbide material using ion amplitude dielectric film process technology, and the scattering mirror is made of coated silicon wafer.

6. The apparatus for the neural network-based ultra-low power optical target detection method according to claim 4, characterized in that: The optical prism device is followed by a beam splitter.

Images