Method and system for quickly identifying multi-source photoelectric detection image based on neural network structure

Abstract

The invention discloses a fast recognition method and system for a multi-source photoelectric detection image based on a neural network structure, and relates to the technical field of image recognition. Acquiring images received by a plurality of photoelectric detectors as input original images; establishing a neural network according to the probe pixels and the classification number of the input original images, performing automatic pruning processing on the neural network, training the pruned neural network based on the input original images, and obtaining neural network weight information and image category information; constructing a bias processor, wherein the bias processor is used for modulating an output signal of the photoelectric detector; and correspondingly connecting the trained neural network with a bias processor to obtain a rapid recognition result of the image. According to the method, the photoelectric detection technology and the deep learning technology are effectively integrated, efficient, accurate and rapid image recognition is achieved, pruning and optimization algorithms in deep learning are combined, the recognition speed and efficiency are improved, and meanwhile resource consumption is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and more particularly to a method and system for quickly recognizing multi-source photoelectric detection images based on a neural network structure. Background Art

[0002] Photoelectric detection technology uses the photoelectric effect to convert the target's optical signal into an electrical signal. It has a wide range of applications in many key areas. With the development of scene requirements and technology, multi-source photoelectric detection systems based on multiple photoelectric detection methods such as visible light, infrared light, and lidar have emerged. The information from these different sensor sources has its own unique physical properties and information advantages. However, the amount of multi-source photoelectric detection image data has exploded, posing a huge challenge to the rapid and effective processing of information.

[0003] Traditional image recognition technology is difficult to efficiently and accurately identify multi-source photoelectric detection images due to the complexity and differences of the data structure and features of each source image. This is mainly because in the feature extraction process for different source images, complex algorithms need to be designed separately, and when integrating multi-source information, information loss or mismatch problems are prone to occur, resulting in slow recognition speed and low accuracy, which cannot meet the strict requirements of real-time and reliability in practical applications.

[0004] Therefore, how to provide a rapid recognition method and system for multi-source photoelectric detection images based on a neural network structure, efficiently process multi-source photoelectric signals, automatically classify and identify targets, and ensure the optimization of computing speed while improving recognition accuracy and efficiency is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0005] In view of this, the present invention provides a method and system for rapid recognition of multi-source photoelectric detection images based on a neural network structure. By effectively integrating photoelectric detection technology and deep learning technology, efficient and accurate image classification and recognition are achieved, while greatly improving the computing speed to meet image recognition needs.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: a rapid recognition method of multi-source photoelectric detection images based on a neural network structure, comprising:

[0007] Obtaining images received by a plurality of photodetectors as input original images;

[0008] According to the probe pixels of the photodetector and the classification number of the input original image, a neural network is constructed, the neural network is automatically pruned, and the pruned neural network is trained based on the input original image to obtain neural network weight information and image category information;

[0009] constructing a bias processor, the bias processor being used to modulate an output signal of the photodetector;

[0010] The trained neural network is connected to the bias processor accordingly to obtain the image recognition result.

[0011] Preferably, the automatic pruning process of the neural network includes:

[0012] Adding a super network after each layer of the unpruned neural network; the super network is a structure consisting of two fully connected layers;

[0013] Cluster the filters of each layer in the unpruned neural network to identify the prunable parts. Based on the clustering results, use a genetic algorithm to search for the optimal sub-pruning rate of the neural network and generate a pruning strategy.

[0014] The neural network is pruned according to the pruning strategy to obtain a pruned neural network.

[0015] Provide a neural network model with shorter computing time and less storage resources for subsequent deep learning tasks.

[0016] Preferably, the method further comprises: after adding the hypernetwork, training the hypernetwork, wherein the training process of the hypernetwork comprises:

[0017] Randomly generate filter weights for each fully connected layer in the neural network;

[0018] constructing a random network based on the filter weights;

[0019] Update the parameters of the hypernetwork to obtain the trained hypernetwork.

[0020] Preferably, the pruning strategy includes:

[0021] Select ω channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels ω selected in each layer of the neural network to the total number of channels β in the layer is equal to the optimal sub-pruning rate, complete the pruning of the neural network, and obtain the optimal sub-network of the neural network.

[0022] Preferably, searching for the optimal sub-pruning rate of the neural network using a genetic algorithm comprises:

[0023] Set the maximum number of iterations of the genetic algorithm and calculate the total number of parameters B of the neural network 0 and the amount of calculation D 0 ;

[0024] Initialize M groups of pruning rates, each group includes N different pruning rates, where N is equal to the number of network layers;

[0025] Encode each group of pruning rates into binary codes respectively, and perform crossover and mutation operations on the binary codes to generate a number of new binary codes, and obtain the sum of the number of newly generated binary codes and the number of original binary codes as P;

[0026] Decode P binary codes and decode each binary code into a set of pruning rates p i,j ;

[0027] According to the pruning rate obtained by decoding, the network model is adjusted to select channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels selected in each layer of the neural network to the total number of channels in the layer is equal to the pruning rate;

[0028] Delete the selected channels from the neural network to obtain P sub-networks, each of which corresponds to a set of pruning rates;

[0029] Calculate the accuracy, parameter amount and computational complexity of each sub-network, and calculate the fitness of each sub-network;

[0030] The probability q of each network being selected is calculated by the fitness of each sub-network i , through the roulette wheel selection method, select R sub-networks from P sub-networks, and obtain the current pruning rate p corresponding to each sub-network i,j , 1≤R≤P;

[0031] When the number of iterations reaches the set maximum number of iterations, the search is completed and the optimal sub-pruning rate of the neural network model corresponding to the optimal fitness is obtained.

[0032] Preferably, a bias processor is designed according to the extracted neural network weights, each weight corresponds to a different photodetector channel, and a corresponding adjustment circuit is designed to adjust the output signal of each channel;

[0033] The regulating circuit includes an amplifier, a filter and a modulator;

[0034] The amplifier is used to enhance the signal strength for better transmission and processing;

[0035] The filter is used to eliminate noise and ensure signal quality;

[0036] The modulator is used to adjust the output characteristics of the corresponding photodetector according to the weight information.

[0037] Preferably, the working principle of the regulating circuit includes:

[0038] Obtain output signals from multiple photodetectors, analyze frequency components in the signals, and calculate power values; based on the power values, obtain bias voltages of several operating points of the modulator;

[0039] The first comparator threshold and the second comparator threshold are set respectively to obtain a threshold range, thereby ensuring high accuracy under different signal conditions.

[0040] extracting a DC component from the output signal, and outputting a level signal based on the DC component;

[0041] It is determined whether the bias voltage corresponding to the level signal is within the threshold range, and based on the determination result, the modulator operating point is stabilized, and automatic adjustment is achieved to ensure the stability of the modulator.

[0042] Preferably, the modulator's several operating points include: a modulator maximum operating point, a modulator minimum operating point, and a modulator linear operating point, to ensure the accuracy of the modulator under different signal conditions.

[0043] Preferably, a rapid recognition system for multi-source photoelectric detection images based on a neural network structure comprises:

[0044] An image acquisition module is used to obtain images received by multiple photoelectric detectors as input original images;

[0045] A neural network construction and training module is used to build a neural network according to the probe pixels of the photodetector and the number of classifications of the input original image, automatically prune the neural network, train the pruned neural network based on the input original image, and obtain neural network weight information and image category information;

[0046] A bias processor building module, used to build a bias processor, wherein the bias processor is used to modulate the output signal of the photodetector;

[0047] The target recognition module is used to connect the trained neural network to the bias processor to obtain the recognition result of the image.

[0048] Through the above technical solutions, it can be known that compared with the prior art, the present invention discloses a method and system for rapid recognition of multi-source photoelectric detection images based on a neural network structure. The present invention obtains images received by multiple photoelectric detectors as input original images; according to the probe pixels of the photoelectric detector and the number of categories of the input original image, a neural network is constructed, and the neural network is automatically pruned. The pruned neural network is trained based on the input original image to obtain neural network weight information and image category information; a bias processor is constructed, and the bias processor is used to modulate the output signal of the photoelectric detector; the trained neural network is connected to the bias processor accordingly to obtain the recognition result of the image. A neural network is constructed according to the number of probe pixels of the photoelectric detector and the number of categories of the input image. The constructed neural network is automatically pruned to reduce unnecessary parameters and calculations, thereby improving the efficiency of the network. The pruned neural network is trained based on the input original image to obtain the weight information of the network and the category information of the image, ensuring that the network can reasonably and accurately classify the target object. Weight information is extracted from the trained neural network, and the bias processor is constructed using this information. The function of the bias processor is to adjust the output signal of the photoelectric detector to adapt to the subsequent recognition process. The present invention can realize efficient image processing and object recognition by combining photoelectric detection technology with neural network, is suitable for scenes with multiple light sources and multiple information sources, and has great application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0050] Figure 1 A schematic flow chart of a method for rapid recognition of multi-source photoelectric detection images based on a neural network structure provided by the present invention.

[0051] Figure 2 A schematic diagram of the structure of a rapid recognition system for multi-source photoelectric detection images based on a neural network structure provided by the present invention.

[0052] Figure 3 A model structure diagram of a neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The embodiment of the present invention discloses a method and system for rapid recognition of multi-source photoelectric detection images based on a neural network structure, which realizes efficient, accurate and rapid image recognition by effectively integrating photoelectric detection technology and deep learning technology, and combines pruning and optimization algorithms in deep learning to improve recognition speed and efficiency while reducing resource consumption.

[0055] In a specific embodiment of the present invention, a fast recognition method for multi-source photoelectric detection images based on a neural network structure is provided. Figure 1 As shown, including:

[0056] The images received by multiple photoelectric detectors are obtained as input original images; the photoelectric detectors include a collection module and a spectral imaging module; the collection module collects reflected light of the object after emitting rays, generates different reflected electrical signals according to the wavelength of the reflected light, and the spectral imaging module generates an image according to the electrical signals;

[0057] According to the probe pixels of the photodetector and the classification number of the input original image, a neural network is constructed, the neural network is automatically pruned, and the pruned neural network is trained based on the input original image to obtain neural network weight information and image category information;

[0058] Constructing a bias processor, the bias processor is used to modulate the output signal of the photodetector; there are multiple bias processors, each bias processor corresponds to a photodetector;

[0059] The trained neural network is connected to the bias processor accordingly to obtain the image recognition result.

[0060] Specifically, Figure 3 As shown, the model structure of the neural network includes: an image input layer, a conv1 convolution layer, a conv2 convolution layer, a conv3 convolution layer, a conv4 convolution layer, a conv5 convolution layer, an fc6 fully connected layer, an fc7 fully connected layer and an output classification layer.

[0061] Specifically, the process of training the pruned neural network based on the input original image to obtain the neural network weight information and image category information is based on the following formula:

[0062]

[0063] Among them, C represents the target category, C i represents the i-th category, n represents the number of image pixels, t m represents the mth pixel, P im represents the weight of the mth pixel of the i-class image, b i represents bias, f is the activation function corresponding to the summing node to the final output value, the activation function does not change the relative size of the output value, but only makes the large value larger and the small value smaller, so as to obtain a clearer and more accurate classification result;

[0064] The trained neural network weights are then extracted and the corresponding bias processor is made in proportion.

[0065] Specifically, the automatic pruning process of the neural network includes:

[0066] Adding a super network after each layer of the unpruned neural network; the super network is a structure consisting of two fully connected layers;

[0067] Cluster the filters of each layer in the unpruned neural network to identify the prunable parts. Based on the clustering results, use a genetic algorithm to search for the optimal sub-pruning rate of the neural network and generate a pruning strategy.

[0068] The neural network is pruned according to the pruning strategy to obtain a pruned neural network.

[0069] Provide a neural network model with shorter computing time and less storage resources for subsequent deep learning tasks.

[0070] Specifically, it also includes: after adding the super network, training the super network, and the training process of the super network includes:

[0071] Randomly generate filter weights for each fully connected layer in the neural network; for each layer of the neural network before pruning, randomly input the one-dimensional vector representing the random network structure into the hypernetwork to obtain the weight matrix, and take out the previously specified number of elements as the weights of the current layer filter of the random network;

[0072] constructing a complete random network based on the filter weights;

[0073] Update the parameters of the hypernetwork to obtain the trained hypernetwork.

[0074] The embodiment of the present invention introduces a super network to quickly obtain the sub-network weights. Compared with the previous method of simply pruning the weights of the unpruned network to obtain the random network weights, the random network weights obtained in the embodiment of the present invention are more reasonable, which will accelerate the convergence speed of pruning using reinforcement learning.

[0075] Specifically, the filters of each layer in the unpruned neural network are clustered, including:

[0076] Map the filters of the neural network into points, randomly form class centers, and initialize the class centers;

[0077] In response to completing the initialization of the plurality of cluster centers, generating a distance matrix;

[0078] Cluster multiple filters at each layer based on the distance matrix;

[0079] In response to completing the clustering of the filters, updating the cluster centers;

[0080] Determine whether the updated class center has changed;

[0081] If there is no change, calculate the silhouette coefficient for each cluster center.

[0082] Among them, the distance d(i,i') from each filter to the center of each class is specifically expressed as:

[0083]

[0084] Among them, n represents the dimension of the high-dimensional space, i and i' represent the two points where the distance is calculated, m represents a natural number, and i m Indicates the position of the filter, i' m Indicates the location of the cluster center.

[0085] Clustering is introduced as an evaluation indicator to measure the importance of filters, that is, filters close to the class center are considered redundant and unimportant, which increases the accuracy of the pruned sub-network. The class center of each layer is the point mapped by k randomly selected filters. Furthermore, clustering is performed on each layer of unpruned filters until the results converge. After completing the clustering of the first layer, the second layer of clustering is performed.

[0086] Specifically, the pruning strategy includes:

[0087] Select ω channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels ω selected in each layer of the neural network to the total number of channels β in the layer is equal to the optimal sub-pruning rate, complete the pruning of the neural network, and obtain the optimal sub-network of the neural network.

[0088] Specifically, a genetic algorithm is used to search for the optimal sub-pruning rate of a neural network, including:

[0089] Set the maximum number of iterations of the genetic algorithm and calculate the total number of parameters B of the neural network 0 and the amount of calculation D 0 ;

[0090] Initialize M groups of pruning rates, each group includes N different pruning rates, where N is equal to the number of network layers;

[0091] Encode each group of pruning rates into binary codes respectively, and perform crossover and mutation operations on the binary codes to generate a number of new binary codes, and obtain the sum of the number of newly generated binary codes and the number of original binary codes as P;

[0092] Decode P binary codes and decode each binary code into a set of pruning rates p i,j ;

[0093] According to the pruning rate obtained by decoding, the network model is adjusted to select channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels selected in each layer of the neural network to the total number of channels in the layer is equal to the pruning rate;

[0094] Delete the selected channels from the neural network to obtain P sub-networks, each of which corresponds to a set of pruning rates;

[0095] Calculate the accuracy of each sub-network a 0 , parameter b 0 and the amount of calculation d 0 , and using a 0 , b 0 and d 0 Calculate the fitness f of each sub-network i ; The formula is as follows:

[0096]

[0097] Among them, a i , b i and d i Respectively represent the accuracy, parameter amount and calculation amount of each sub-network; θ is the hyperparameter of each part of the balance formula, -1≤θ≤1, ε represents the allowable deviation of accuracy;

[0098] Through the fitness of each sub-network f i Calculate the probability q of each network being selected i , through the roulette wheel selection method, select R sub-networks from P sub-networks, and obtain the current pruning rate p corresponding to each sub-network i,j , 1≤R≤P;

[0099] When the number of iterations reaches the set maximum number of iterations, the search is completed and the optimal sub-pruning rate p of the neural network model corresponding to the optimal fitness is obtained. best .

[0100] The embodiment of the present invention combines pruning and optimization algorithms in deep learning to improve recognition speed and efficiency while reducing resource consumption. Genetic algorithms are used to perform heuristic search for pruning rates suitable for neural networks, and the search direction is changed by changing the calculation method of fitness, thereby changing the final search results, and adjusting pruning according to actual needs.

[0101] The pruned neural network is trained based on the input image and the original image to obtain the neural network weight information and image category information, including:

[0102] Collect a large number of input images and their corresponding original images as training data sets, and preprocess these images, including but not limited to image normalization, resizing, etc., to ensure the consistency and validity of the input data;

[0103] Feed the preprocessed input image and the original image into the pruned neural network;

[0104] Use optimization algorithms to continuously adjust the weights of the neural network so that the network output is as close to the actual image category information as possible;

[0105] During the training process, pay close attention to the loss function value of the network. The loss function can measure the difference between the network output and the true label. The network performance can be optimized by continuously reducing the loss function value.

[0106] Use a validation set to test the network's performance on unseen data to prevent overfitting;

[0107] Based on the feedback from the training data, the weight information of the neural network is continuously updated, and for each layer of neurons, the weight value is adjusted according to its contribution to the final output.

[0108] Specifically, a bias processor is designed according to the extracted neural network weights, each weight corresponds to a different photodetector channel, and a corresponding adjustment circuit is designed to adjust the output signal of each channel;

[0109] The regulating circuit includes an amplifier, a filter and a modulator;

[0110] The amplifier is used to enhance the signal strength for better transmission and processing;

[0111] The filter is used to eliminate noise and ensure signal quality;

[0112] The modulator is used to adjust the output characteristics of the corresponding photodetector according to the weight information.

[0113] Implement the logic combined with the neural network reasoning process to adjust the voltage of the photodetector according to the weight information. A PID controller or other adaptive control algorithm can be used here to optimize the output.

[0114] At the same time, a feedback system is designed to monitor the characteristics of the output signal (such as amplitude, frequency characteristics, etc.) and feed back to the bias processor to further adjust the voltage setting to ensure that the output signal meets expectations.

[0115] Ensure that the bias processor can be effectively connected to the photodetector and neural network for real-time data transmission and processing to respond to different environmental changes and recognition tasks.

[0116] Specifically, the working principle of the regulating circuit includes:

[0117] Obtain output signals from multiple photodetectors, analyze frequency components in the signals, and calculate power values; based on the power values, obtain bias voltages of several operating points of the modulator;

[0118] The first comparator threshold and the second comparator threshold are set respectively to obtain a threshold range, thereby ensuring high accuracy under different signal conditions.

[0119] extracting a DC component from the output signal, and outputting a level signal based on the DC component;

[0120] It is determined whether the bias voltage corresponding to the level signal is within the threshold range, and based on the determination result, the modulator operating point is stabilized, and automatic adjustment is achieved to ensure the stability of the modulator.

[0121] Specifically, the process of outputting the level signal includes:

[0122] Based on the DC quantity output by the mixer, a DC component is obtained through processing by a power divider. The DC component passes through a first comparator and a second comparator respectively to output a high level signal or a low level signal.

[0123] Specifically, the process of detecting and judging the bias voltage corresponding to the level signal includes:

[0124] If the bias voltage is within the threshold range, the voltage is output through the digital-to-analog converter and fed back to the bias voltage control port of the modulator; otherwise, the bias voltage of the modulator continues to be adjusted.

[0125] The embodiment of the present invention realizes automatic bias control of the photodetector and locks the optimal working point of the modulator through a combination of software and hardware. The present invention has a reasonable design, is easy to implement and integrate, shortens the time of modulator bias control, has a compact structure, improves operability, reduces the number of components, and saves costs.

[0126] Specifically, the modulator's several operating points include: a modulator maximum operating point, a modulator minimum operating point, and a modulator linear operating point, to ensure the accuracy of the modulator under different signal conditions.

[0127] In a specific embodiment of the present invention, the method of the embodiment of the present invention is applied to an intelligent security monitoring system. In a large warehouse security scene, multiple different types of photoelectric detectors are arranged, including visible light cameras, infrared thermal imagers, etc., for all-round monitoring of the situation in the warehouse. These photoelectric detectors will continuously receive image information from all directions. The method of the embodiment of the present invention can quickly identify these multi-source photoelectric detection images to detect whether there are abnormal situations, such as human intrusion, fire hazards, etc.

[0128] Specifically, multiple photoelectric detectors (visible light cameras, infrared thermal imagers, etc.) collect images in the warehouse in real time, and these images are transmitted to the system as input images and original images. Among them, the visible light camera captures visible light information such as the appearance of objects and personnel activities in the warehouse, and the infrared thermal imager detects the thermal radiation distribution of objects and obtains the temperature characteristic image of the object surface.

[0129] The probe pixels of the visible light camera are set to [X1, Y1], and the probe pixels of the infrared thermal imager are set to [X2, Y2]. Considering these pixel information and the number of image categories that need to be identified (the image category is represented by Z), the number of layers and nodes of the neural network are designed. After the convolution layer of the neural network, a super network structure is added to quickly obtain the sub-network weights later.

[0130] The pruned neural network is trained based on the input image and the original image. A large amount of labeled image data (image samples labeled with categories such as people, goods, equipment, etc.) is used to train the neural network.

[0131] During the training process, the weights of the neural network are continuously adjusted to make the network output as close to the real image category information as possible. The loss function value of the network is continuously reduced through optimization algorithms (such as stochastic gradient descent, etc.), and the network is regularly evaluated using the validation set to prevent overfitting. After multiple iterations of training, the neural network gradually learns the feature representation of the input image and accurately predicts the category information of the image, thereby obtaining the neural network weight information and image category information.

[0132] According to the neural network weight information extracted from the trained neural network, a bias processor is designed, each weight corresponds to a different photodetector channel, and a corresponding adjustment circuit is designed for each channel to adjust the output signal. In an embodiment of the present invention, for a channel of a visible light camera, corresponding adjustment circuit parameters are designed according to its corresponding weight.

[0133] The bias processor is used to adjust the voltage of the photodetector in different ways. According to different scene requirements and the characteristics of the photodetector, the bias processor adjusts the voltage so that the photodetector can better capture the characteristic information of the target object. In an embodiment of the present invention, when a human intrusion is detected, the bias processor adjusts the bias of the infrared thermal imager so that it can more keenly detect the thermal radiation characteristics of the human body.

[0134] The bias-adjusted photodetector signals are connected according to the corresponding relationship in the trained neural network. The neural network has learned the associations and characteristic patterns between different photodetector signals. Through this connection method, the information of multi-source photodetection images can be integrated. The recognition result of the target object is obtained according to the output of the neural network. In an embodiment of the present invention, when a person enters the warehouse, the system can accurately identify the person's location, behavior and other information, and issue an alarm in time. At the same time, abnormal situations such as fire hazards can also be quickly and accurately identified through comprehensive analysis of multi-source photodetection images.

[0135] In a specific embodiment of the present invention, a fast recognition system for multi-source photoelectric detection images based on a neural network structure is provided. Figure 2 As shown, including:

[0136] An image acquisition module is used to obtain images received by multiple photoelectric detectors as input original images;

[0137] A neural network construction and training module is used to build a neural network according to the probe pixels of the photodetector and the number of classifications of the input original image, automatically prune the neural network, train the pruned neural network based on the input original image, and obtain neural network weight information and image category information;

[0138] A bias processor building module, used to build a bias processor, wherein the bias processor is used to modulate the output signal of the photodetector;

[0139] The target recognition module is used to connect the trained neural network to the bias processor to obtain the recognition result of the image.

[0140] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0141] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A fast recognition method for multi-source photoelectric detection images based on a neural network structure, characterized in that: include: Obtaining images received by a plurality of photodetectors as input original images; According to the probe pixels of the photodetector and the classification number of the input original image, a neural network is constructed, the neural network is automatically pruned, and the pruned neural network is trained based on the input original image to obtain neural network weight information and image category information; constructing a bias processor, the bias processor being used to modulate an output signal of the photodetector; The trained neural network is connected to the bias processor accordingly to obtain the image recognition result.

2. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 1 is characterized in that: The neural network is automatically pruned, including: Add a hypernetwork after each layer of the unpruned neural network; Cluster the filters of each layer in the unpruned neural network, and based on the clustering results, use a genetic algorithm to search for the optimal sub-pruning rate of the neural network and generate a pruning strategy; The neural network is pruned according to the pruning strategy to obtain a pruned neural network.

3. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 2 is characterized in that: Also includes: After adding the hypernetwork, the hypernetwork is trained. The training process of the hypernetwork includes: Randomly generate filter weights for each fully connected layer in the neural network; constructing a random network based on the filter weights; Update the parameters of the hypernetwork to obtain the trained hypernetwork.

4. The rapid recognition method of multi-source photoelectric detection images based on a neural network structure according to claim 2 is characterized in that: The pruning strategies include: Select ω channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels ω selected in each layer of the neural network to the total number of channels β in the layer is equal to the optimal sub-pruning rate, complete the pruning of the neural network, and obtain the optimal sub-network of the neural network.

5. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 4 is characterized in that: Genetic algorithm is used to search for the best sub-pruning rate of neural network, including: Set the maximum number of iterations of the genetic algorithm and calculate the total parameter amount B0 and calculation amount D0 of the neural network; Initialize M groups of pruning rates, each group includes N different pruning rates, N is equal to the number of network layers; Encode each group of pruning rates into binary codes respectively, and perform crossover and mutation operations on the binary codes to generate a number of new binary codes, and obtain the sum of the number of newly generated binary codes and the number of original binary codes as P; Decode P binary codes and decode each binary code into a set of pruning rates p i,j ; According to the pruning rate obtained by decoding, the network model is adjusted to select channels with smaller scaling factor values ​​in each layer of the neural network so that the ratio of the number of channels selected in each layer of the neural network to the total number of channels in the layer is equal to the pruning rate; Delete the selected channels from the neural network to obtain P sub-networks, each of which corresponds to a set of pruning rates; Calculate the accuracy, parameter amount, and computational complexity of each sub-network, and calculate the fitness of each sub-network; The probability of each network being selected is calculated through the fitness of each sub-network. R sub-networks are selected from P sub-networks through the roulette wheel selection method to obtain the current pruning rate p corresponding to each sub-network. i,j , 1≤R≤P; When the number of iterations reaches the set maximum number of iterations, the search is completed and the optimal sub-pruning rate of the neural network model corresponding to the optimal fitness is obtained.

6. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 1 is characterized in that: Design a bias processor based on the extracted neural network weights, each weight corresponds to a different photodetector channel, and design a corresponding adjustment circuit to adjust the output signal of each channel; The regulating circuit includes an amplifier, a filter and a modulator; The amplifier is used to enhance signal strength; The filter is used to eliminate noise; The modulator is used to adjust the output characteristics of the corresponding photodetector according to the weight information.

7. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 6 is characterized in that: The working principle of the regulating circuit includes: Obtain output signals from multiple photodetectors, analyze frequency components in the signals, and calculate power values; based on the power values, obtain bias voltages of several operating points of the modulator; respectively setting a first comparator threshold and a second comparator threshold to obtain a threshold interval; extracting a DC component from the output signal, and outputting a level signal based on the DC component; It is determined whether the bias voltage corresponding to the level signal is within the threshold range, and based on the determination result, the modulator operating point is stabilized.

8. The method for rapid recognition of multi-source photoelectric detection images based on a neural network structure according to claim 7 is characterized in that: The several working points of the modulator include: a maximum working point of the modulator, a minimum working point of the modulator and a linear working point of the modulator.

9. A fast recognition system for multi-source photoelectric detection images based on a neural network structure, characterized in that: include: An image acquisition module is used to obtain images received by multiple photoelectric detectors as input original images; A neural network construction and training module is used to build a neural network according to the probe pixels of the photodetector and the number of classifications of the input original image, automatically prune the neural network, train the pruned neural network based on the input original image, and obtain neural network weight information and image category information; A bias processor building module, used to build a bias processor, wherein the bias processor is used to modulate the output signal of the photodetector; The target recognition module is used to connect the trained neural network to the bias processor to obtain the recognition result of the image.

Images