Non-cooperative target component intelligent perception method, electronic device and storage medium

Abstract

The present invention discloses a method for intelligently perceiving non-cooperative target components, an electronic device and a storage medium. The method is used for dynamically performing online intelligent recognition and detection of non-cooperative target components in space, including: constructing a pre-trained feature extraction network model; combining the pre-trained feature extraction network model, constructing a target rapid detection network model based on a small sample of meta-learning; inputting the image data set to be detected of the non-cooperative target components acquired in real time into the target rapid detection network model, and outputting the perception result. The present invention solves the problem of dynamic online intelligent recognition and detection of non-cooperative target components in tasks such as on-orbit capture and fine operation of component disassembly by a robotic arm, and has the advantages of high generalization of target recognition, fast learning speed and strong environmental adaptability.

Description

Technical Field

[0001] The present invention relates to the field of space attack and defense technology, and in particular to a non-cooperative target component intelligent perception method, electronic equipment and storage medium for robot-assisted real-time online incremental learning. Background Art

[0002] In order to achieve typical on-orbit service and control tasks such as flexible capture of spacecraft, module replacement, and component disassembly, it is necessary to solve the identification, measurement, and 3D reconstruction technology of the operating parts of failed non-cooperative targets. With the rapid development of aerospace technology, future autonomous on-orbit services will show a clear trend of more unknown and changeable target characteristics and more diverse and sophisticated operating tasks. The traditional non-cooperative target batch learning recognition algorithm has the problems of weak generalization for unknown targets and diverse operating tasks, low learning efficiency, and poor adaptability to incremental data models. Summary of the invention

[0003] The purpose of the present invention is to provide a non-cooperative target component intelligent perception method, electronic device and storage medium for robot-assisted real-time online incremental learning, so as to solve the problems of weak generalization for unknown targets and diverse operation tasks, low learning efficiency and poor adaptability to incremental data models existing in traditional non-cooperative target batch learning and recognition algorithms.

[0004] In order to solve the above problems, the present invention is implemented by the following technical solutions:

[0005] A non-cooperative target component intelligent perception method is used for dynamically performing online intelligent recognition and detection of spatial non-cooperative target components, comprising: step S1, constructing a pre-trained feature extraction network model. Step S2, combining the pre-trained feature extraction network model, and constructing a target rapid detection network model based on a small sample of meta-learning. Step S3, inputting the image data set to be detected of the non-cooperative target component acquired in real time into the target rapid detection network model, and outputting the perception result.

[0006] Optionally, the step S1 includes: step S11, obtaining a labeled natural image dataset, using the labeled natural image dataset to train a first convolutional neural network to obtain weight parameters of a feature extraction basic network model.

[0007] Step S12: Obtain a labeled satellite image dataset.

[0008] Step S13: using the weight parameters of the feature extraction basic network model as the initial values ​​of the second convolutional neural network model, optimizing the second convolutional neural network model by iteratively adjusting the weight parameters, inputting the labeled satellite image data set into the optimized second convolutional neural network for training until effective image features are obtained, thereby obtaining the pre-trained feature extraction network model.

[0009] Optionally, the step S11 includes: the labeled natural image dataset is a preset open source large-scale labeled natural image dataset;

[0010] The step S12 includes: acquiring and screening a plurality of space non-cooperative target images under the target satellite's on-orbit service and control mission, and labeling the screened space non-cooperative target images to obtain the labeled satellite image data set.

[0011] Optionally, step S12 includes: the space non-cooperative target images included in the labeled satellite image dataset include satellite images of rotating moving targets of multiple types and multiple perspectives under different missions, different illuminations, different backgrounds, and different scales. The labeled satellite image dataset includes space non-cooperative target images under multiple complex and harsh environments coupled with local overexposure, feature loss, different scales, image blur, and target rotation.

[0012] Optionally, screening a plurality of the spatial non-cooperative target images includes the following process:

[0013] The collected original plurality of space non-cooperative target images are preprocessed by image global contrast normalization. The preprocessed plurality of space non-cooperative target images are image scaled and data augmented based on optimized classification to effectively expand and screen the plurality of space non-cooperative target images, thereby obtaining the labeled satellite image data set.

[0014] The global contrast normalization preprocessing of the collected original several spatial non-cooperative target images includes the following process: Assume that the original spatial non-cooperative target image is represented by a tensor:

[0015]

[0016] Where X represents the intensity of the image; represents the tensor representation symbol; m represents the image row; n represents the image column;

[0017] The contrast of the entire image is expressed as follows:

[0018]

[0019] In the formula, the value range of k is 1 to 3. When k = 1, X i,j,1 represents the intensity of red in the i-th row and j-th column; when k = 2, X i,j,2 represents the intensity of green in row i and column j; when k = 3, X i,j,3 represents the intensity of blue in row i and column j;

[0020] is the average intensity of the entire image, satisfying:

[0021]

[0022] Image generated by global contrast normalization It is expressed by the following formula:

[0023]

[0024] Where ε = 10 -9 , λ=0, s=1; the data augmentation based on optimized classification first calculates the classification accuracy of each category in the test set, and then finds the category with the lowest accuracy in the test set.

[0025] Optionally, the step S2 includes: step S21, using the robot's end hand-eye camera to dynamically capture images of non-cooperative target components, randomly screening a preset amount of the images for data preprocessing, and annotating the screened images to obtain sample labels, and obtain a small sample training data set of non-cooperative target components under the current operation task.

[0026] Step S22: using the weight parameters of the pre-trained feature extraction network model as initialization parameters of the feature extraction network model, and using a meta-learning method in combination with the small sample training data set to fine-tune the feature extraction network model to obtain a lightweight backbone network model.

[0027] Step S23: adopt a convex optimization method to train the lightweight backbone network model to obtain a target rapid detection network model.

[0028] Optionally, step S21 includes: the step of dynamically collecting images of non-cooperative target components using the end hand-eye camera of the robot includes: in a robot-assisted non-cooperative target dynamic capture test system in a ground simulation environment, by controlling the end of the robot's mechanical arm to move from far to near, and using the end hand-eye camera of the robot to dynamically collect image data of non-cooperative target component components of the current operation task in the 2m-0.2m approach process. The number of image data of each component in the non-cooperative target component assembly is 100.

[0029] Optionally, the step S3 includes: step S31, using the target rapid detection network model to quickly train the image data stream acquired in real time by the end of the robot's mechanical arm, and obtaining the state information parameters of the non-cooperative target component in real time. Step S32, online annotation correction of images with a confidence level lower than 95% in the state information parameters and re-optimization and training of the parameters of the target rapid detection network model. Step S33, using online incremental learning optimization based on super-participation weight optimization knowledge distillation joint loss under time constraints to train the target rapid detection network model, and real-time update and estimate of the model parameters of the sample stream data acquired in real time for the current task, one by one, trained and fed back.

[0030] On the other hand, the present invention further provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method described above is implemented.

[0031] In yet another aspect, the present invention further provides a readable storage medium, wherein the readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0032] The present invention has at least one of the following advantages:

[0033] In order to realize the autonomy and intelligence of typical on-orbit service and manipulation tasks in space, the present invention develops space non-cooperative target perception technology of robot-assisted visual information, and realizes generalized intelligent perception of non-cooperative target capture parts and component disassembly parts through a real-time online self-learning system, effectively improving the system's adaptability to complex space environments, unknown dynamic targets, and various fine operations, and improving the generalized intelligent recognition performance of multi-task space non-cooperative targets, providing technical support for the on-orbit capture and component disassembly fine operation tasks of the robotic arm.

[0034] Compared with the traditional offline batch learning recognition and detection algorithm for non-cooperative targets, the method of the present invention adopts an online incremental self-learning algorithm based on real-time perception information to realize the processing of streaming data and complete the generalized intelligent perception of the capture parts and component disassembly parts of non-cooperative targets. It can effectively improve the system's adaptability to complex spatial environments, unknown dynamic targets, and diverse fine operations, and improve the recognition performance of multi-task spatial non-cooperative targets. It has the advantages of high generalization of target recognition, fast learning speed, and strong environmental adaptability, and can comprehensively improve the level of intelligent autonomy in space operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A flow chart of a non-cooperative target component intelligent perception method provided by one embodiment of the present invention;

[0036] Figure 2 A flowchart for implementing step S1 in a non-cooperative target component intelligent perception method provided by one embodiment of the present invention;

[0037] Figure 3 A flowchart for implementing step S2 in a non-cooperative target component intelligent perception method provided by one embodiment of the present invention;

[0038] Figure 4 This is a flowchart for implementing step S3 in a non-cooperative target component intelligent perception method provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following is a further detailed description of a robot-assisted real-time online incremental learning non-cooperative target component intelligent perception method, electronic device and storage medium proposed in the present invention in combination with the accompanying drawings and specific embodiments. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purposes, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present invention.

[0040] like Figure 1 As shown, the present embodiment provides a non-cooperative target component intelligent perception method, which is used to dynamically perform online intelligent recognition and detection of spatial non-cooperative target components, including:

[0041] Step S1: construct a pre-trained feature extraction network model.

[0042] Step S2: Combine the pre-trained feature extraction network model and build a target rapid detection network model based on a small sample of meta-learning.

[0043] Step S3, input the image data set to be detected of the non-cooperative target component acquired in real time into the target rapid detection network model, and output the perception result. That is, the dynamic recognition and detection of the non-cooperative target component by robot-assisted real-time online incremental learning is realized, and the dynamic online intelligent perception of the non-cooperative target capture part and component disassembly part in the image acquired in real time at the end of the robot arm is realized.

[0044] like Figure 2As shown, the step S1 of this embodiment includes:

[0045] Step S11, obtaining a labeled natural image data set; using the labeled natural image data set to train a first convolutional neural network to obtain weight parameters (initial parameter values) of a feature extraction basic network model.

[0046] The step S11 specifically includes: the labeled natural image dataset is a preset open source ImagNet natural large-scale labeled image dataset.

[0047] Step S12, obtaining a labeled satellite image dataset. In this embodiment, step S12 includes: acquiring and screening a number of space non-cooperative target images under the target satellite's on-orbit service and control mission, and annotating the screened space non-cooperative target images to obtain the labeled satellite image dataset.

[0048] Specifically, step S12 collects and screens space non-cooperative target images under typical on-orbit service and control tasks such as flexible capture of aircraft, module replacement, and component disassembly, and annotates the screened 5,000 images (space non-cooperative target images) to obtain sample labels, thereby completing the construction of a labeled training data set for key components of 20 types of space non-cooperative targets (labeled satellite image data set).

[0049] The space non-cooperative target images contained in the labeled satellite image data set include satellite images of rotating moving targets of multiple types and multiple perspectives under different missions, different illuminations, different backgrounds, and different scale conditions.

[0050] The labeled satellite image dataset contains spatial non-cooperative target images under a variety of complex and harsh environments coupled with local overexposure, feature loss, different scales, image blur and target rotation.

[0051] The step S12 further includes: screening a plurality of the spatial non-cooperative target images includes the following process:

[0052] The global contrast normalization preprocessing is performed on several original spatial non-cooperative target images collected.

[0053] The pre-processed plurality of space non-cooperative target images are subjected to image scaling and data augmentation based on optimized classification to effectively expand and screen the plurality of space non-cooperative target images, thereby obtaining the labeled satellite image data set.

[0054] The global contrast normalization preprocessing of the collected original several spatial non-cooperative target images includes the following process:

[0055] Assume that the original spatial non-cooperative target image is represented by a tensor:

[0056]

[0057] Where X represents the intensity of the image; represents the tensor representation symbol; m represents the image row; n represents the image column;

[0058] The contrast of the entire image is expressed as follows:

[0059]

[0060] In the formula, the value range of k is 1 to 3. When k = 1, X i,j,1 represents the intensity of red in the i-th row and j-th column; when k = 2, X i,j,2 represents the intensity of green in row i and column j; when k = 3, X i,j,3 Represents the intensity of blue in row i and column j.

[0061] is the average intensity of the entire image, satisfying:

[0062]

[0063] Image generated by global contrast normalization It is expressed by the following formula:

[0064]

[0065] Where ε = 10 -9 ,λ=0,s=1.

[0066] The data augmentation based on optimized classification first calculates the classification accuracy of each category in the test set, and then finds the category with the lowest accuracy in the test set, that is, the classification effect of the feature extraction basic network model for this category is not good, and finally the data augmentation process is performed on this category.

[0067] The 20 types of space non-cooperative target components (space non-cooperative target key components) include pot antennas, thrusters, earth observation cameras, solar panels, solar sensors, infrared horizon instruments, rigid arms, feed sources, flexible arms, solar panel brackets, conical spiral antennas, satellite bodies, star sensors, satellite-rocket docking rings, docking surfaces, laser data transmission, nozzles, measurement and control antennas, end effectors, and GPS antennas, and labelImg is used to annotate images.

[0068] In summary, the global contrast normalization is performed by subtracting the mean value from each image and then rescaling the values ​​to compress them into the interval [0, 1] so that the standard deviation of the pixels is equal to a constant to ensure that the contrast of the image remains unchanged.

[0069] Step S13: using the weight parameters of the feature extraction basic network model as the initial values ​​of the second convolutional neural network model, optimizing and training the model parameters of the second convolutional neural network by iteratively adjusting the weight parameters, inputting the labeled satellite image data set into the optimized second convolutional neural network for training until effective image features are obtained, thereby obtaining the pre-trained feature extraction network model.

[0070] In this embodiment, if Figure 3 As shown, step S2 includes:

[0071] Step S21, using the hand-eye camera at the end of the robot to dynamically collect images of non-cooperative target components, randomly screening a preset amount of the images for data preprocessing, and annotating the screened images to obtain sample labels, thereby obtaining a small sample training data set of non-cooperative target components under the current operation task.

[0072] Specifically,

[0073] The step of dynamically collecting images of non-cooperative target components using a hand-eye camera at the end of the robot comprises:

[0074] In a robot-assisted non-cooperative target dynamic capture test system in a ground simulation environment, the end of the robot's mechanical arm is controlled to move from far to near, and the end hand-eye camera of the robot is used to dynamically collect image data of non-cooperative target component components of the current operation task during the 2m-0.2m approach process.

[0075] The number of image data of each component in the non-cooperative target component assembly is 100, but the present invention is not limited thereto.

[0076] It can be understood that the constructed target component small sample training data set does not exist in the training data set constructed in step S12. This data belongs to small sample data, and the image data of a single component is 100.

[0077] Step S22: using the weight parameters of the pre-trained feature extraction network model as the initialization parameters of the S1 feature extraction network model (DarkNet-53 feature extraction network model), and using a meta-learning method in combination with the small sample training data set to fine-tune the feature extraction network model to obtain a lightweight backbone network model.

[0078] Step S23: train the lightweight backbone network model using a convex optimization method to obtain a target rapid detection network model to achieve rapid identification of different target components. The convex optimization method is to find a point sequence that continuously reduces the objective function value until the trigger stop condition reaches the minimum value.

[0079] Specifically, the step S22 includes: the solution process of the convex optimization method is as follows:

[0080] Let x k is the value of the kth iteration, d k is the k-th search direction, λ k is the step size of the kth iteration, then the kth iteration formula is as follows:

[0081] x k+1 =x k +λ k d k (5)

[0082] In the formula, x k+1 is the value of the k+1th iteration;

[0083] Among them, the k-th search direction satisfies:

[0084]

[0085] f(x k +λ k d k )<f(x k ) (7)

[0086] In the formula, f(x k ) represents the lightweight backbone network model result after the kth iteration training; f(x k +λ k d k ) represents the result of the lightweight backbone network model after the k+1th iteration training.

[0087] It can be seen that here, the feature extraction network model is first fine-tuned by the meta-learning method to obtain the lightweight backbone network model, and then the parameter set θ of the newly learned lightweight backbone network model is optimized by the convex optimization algorithm to update the weights of the obtained parameter set θ so that it can quickly reach the optimal point in the convex function. The optimal parameter solutions θ1, θ2, ...θ for different tasks are achieved in the parameter space. n ,θ n represents the parameter set of the lightweight backbone network model corresponding to the nth target recognition task, and n represents the number of different target recognition tasks. The obtained target fast detection network model has the advantages of small network model and fast processing speed.

[0088] The target recognition rate of the target rapid detection network model is expressed by the following formula using different target recognition average precisions (meanPrecision):

[0089]

[0090] Among them, m is the number of frames, p is i is the recognition rate of each frame image, N is the total number of categories, n pre ∩n gt It indicates that when the predicted value and the true value belong to the same category, the value is 1, and when they are different categories, the value is 0.

[0091] For example, if the true category value of a frame image is [1 1 0 1 1 0 1···n], the category prediction value is [1 11 1 1 0 1···n], and t categories are predicted correctly, then the recognition rate of this frame image is t / n, where n represents the number of categories in the non-cooperative target part components in each frame image.

[0092] The target detection accuracy of the target fast detection network model is expressed by mIoU (mean Intersection-over-Union), as shown in the following formula:

[0093]

[0094] Where M is the number of component categories, n pre ∩n gt Indicates the number of targets whose prediction and true value are of the same category, n pre ∪n gt It indicates the number of targets where both the prediction and the true value have the same category.

[0095] like Figure 4 As shown, step S3 includes:

[0096] Step S31: using the end of the robot's mechanical arm to acquire the image data stream of the key components of the non-cooperative target in real time.

[0097] Step S32: input the image data stream into the target rapid detection network model, and proceed to step S33.

[0098] Step S33: outputting target perception information in real time. The target perception information is state information parameters of the non-cooperative target component obtained in real time; the state information parameters include the type, position and confidence of the non-cooperative target component.

[0099] Step S34, determine whether the confidence in the state information parameter is greater than 95%, if so, proceed to step S35, if not, perform online annotation correction on the image and return to step S32.

[0100] Step S35, use online incremental learning optimization based on super-participation weight optimization knowledge distillation joint loss under time constraints to train the target rapid detection network model and output the perception result; and update and estimate the model parameters of the sample stream data obtained in real time for the current task one by one, and return to step S32.

[0101] The step S35 solves the problems existing in the traditional offline batch learning recognition algorithm, such as weak generalization for unknown targets and diverse operation tasks, low learning efficiency, and poor adaptability to incremental data models. It has the advantages of high generalization for target recognition, fast learning speed, and strong environmental adaptability.

[0102] The knowledge distillation joint loss online incremental learning based on automatic super-weight optimization constructs a new loss function L mlfcnn (p,k * ,t,t * ), so that it can balance the prediction of new classes and the prediction of old classes, thereby minimizing the difference between the response of the original target fast detection network model and the updated target fast detection network model to the old classes, which is specifically implemented as follows:

[0103] Obtain the loss function L of the small sample target fast detection network model based on meta-learning mlfcnn (p,k * ,t,t * ):

[0104]

[0105] Where p is the network response set of all classes, k * is a basic class, t is the output of the localization box regression layer, t * It is the candidate box of the real positioning box; Indicates the current network class; R(tt * ) represents the error value between the current positioning box regression and the true positioning candidate box.

[0106] Calculate the knowledge distillation loss function L dist :

[0107]

[0108] Among them, y A and B Represent the probability output of the new model and the old model for the old class respectively. A and tB is the output of the new model and the old model for the regression position. N is the number of new classes sampled for distillation (the categories of non-cooperative target components currently used for target recognition tasks), |C A | is the number of old target classes (representing the classes of non-cooperative target components that have been originally used for training).

[0109] The total loss L of the online incremental learning training model is the weighted sum of the knowledge distillation loss and the meta-learning-based small sample target fast detection network model loss:

[0110] L=L mlfcnn +λ(t)L dist (12)

[0111] The hyperparameter λ(t) is used to balance the two losses.

[0112] Specifically, λ(t) is obtained by automatic optimization based on time constraints, and the objective function is optimized for target recognition time constraints of different spatial tasks to achieve adaptive dynamic adjustment of hyperparameters.

[0113] In the online incremental learning of knowledge distillation joint loss based on automatic super-weight optimization, the hyperparameter λ(t) of smoothing the "soft label" is set to λ(t<5min), the optimizer is Adam, the epoch of model iteration is 80, the learning rate is 0.001, the learning rate decays by 0.0001 every 20 epochs, and the batch size is 32.

[0114] The learning rate of the target rapid detection network model is 0.0001, the optimizer is Adam, the weight decay rate is 0.00005 after 3K iterations, the momentum (learning decay rate) is set to 0.9, the training time of a single target model is better than 5 minutes, the target recognition rate is better than 95%, and the target detection accuracy is better than 90%.

[0115] On the other hand, the present invention further provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method described above is implemented.

[0116] In yet another aspect, the present invention further provides a readable storage medium, wherein the readable storage medium stores a computer program, and when the computer program is executed by a processor, the method described above is implemented.

[0117] In order to realize the autonomy and intelligence of typical on-orbit service and control tasks in space, this embodiment develops space non-cooperative target perception technology with robot-assisted visual information, and realizes generalized intelligent perception of non-cooperative target capture parts and component disassembly parts through a real-time online self-learning system, effectively improving the system's adaptability to complex space environments, unknown dynamic targets, and diverse fine operations, and improving the generalized intelligent recognition performance of multi-task space non-cooperative targets, providing technical support for the on-orbit capture and component disassembly fine operation tasks of the robotic arm.

[0118] Compared with the traditional offline batch learning recognition and detection algorithm for non-cooperative targets, this embodiment adopts an online incremental self-learning algorithm based on real-time perception information to realize the processing of streaming data and complete the generalized intelligent perception of the capture parts and component disassembly parts of non-cooperative targets. It can effectively improve the system's adaptability to complex spatial environments, unknown dynamic targets, and diverse and fine operations, and improve the recognition performance of multi-task spatial non-cooperative targets. It has the advantages of high generalization of target recognition, fast learning speed, and strong environmental adaptability, and can comprehensively improve the level of intelligent autonomy in space operations.

[0119] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0120] It should be noted that the devices and methods disclosed in the embodiments of this article can also be implemented in other ways. The device implementation described above is only schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of this article. In this regard, each box in the flowchart or block diagram can represent a part of a module, program or code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented by a dedicated hardware-based system for performing a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0121] In addition, the functional modules in the various embodiments of this document may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0122] Although the content of the present invention has been described in detail through the above preferred embodiments, it should be appreciated that the above description should not be considered as a limitation of the present invention. After reading the above content, it will be apparent to those skilled in the art that various modifications and substitutions of the present invention will occur. Therefore, the protection scope of the present invention should be limited by the appended claims.

Claims

1. A non-cooperative target component intelligent perception method for dynamically performing online intelligent recognition and detection of non-cooperative target components in space, characterized in that: include: Step S1, constructing a pre-trained feature extraction network model; Step S2: combining the pre-trained feature extraction network model, and constructing a target rapid detection network model based on a small sample of meta-learning; The step S2 comprises: Step S21: dynamically collect images of non-cooperative target components using the hand-eye camera at the end of the robot, randomly select a preset amount of the images for data preprocessing, and annotate the selected images to obtain sample labels, thereby obtaining a small sample training data set of non-cooperative target components under the current operation task; Step S22: using the weight parameters of the pre-trained feature extraction network model as initialization parameters of the feature extraction network model, and fine-tuning the feature extraction network model using a meta-learning method in combination with the small sample training data set to obtain a lightweight backbone network model; Step S23: training the lightweight backbone network model using a convex optimization method to obtain a target rapid detection network model; Step S3: input the to-be-detected image data set of the non-cooperative target component acquired in real time into the target rapid detection network model, and output the perception result.

2. The non-cooperative target component intelligent perception method according to claim 1, characterized in that: The step S1 comprises: Step S11, obtaining a labeled natural image data set, and using the labeled natural image data set to train a first convolutional neural network to obtain weight parameters of a feature extraction basic network model; Step S12: obtaining a labeled satellite image dataset; Step S13: using the weight parameters of the feature extraction basic network model as the initial values ​​of the second convolutional neural network model, optimizing the second convolutional neural network model by iteratively adjusting the weight parameters, inputting the labeled satellite image data set into the optimized second convolutional neural network for training until effective image features are obtained, thereby obtaining the pre-trained feature extraction network model.

3. The non-cooperative target component intelligent perception method according to claim 2, characterized in that: The step S11 comprises: the labeled natural image dataset is a preset open source large-scale labeled natural image dataset; The step S12 includes: acquiring and screening a plurality of space non-cooperative target images under the target satellite's on-orbit service and control mission, and labeling the screened space non-cooperative target images to obtain the labeled satellite image data set.

4. The non-cooperative target component intelligent perception method according to claim 3, characterized in that: The step S12 includes: the space non-cooperative target images contained in the labeled satellite image data set include rotating moving target satellite images of multiple types and multiple perspectives under different tasks, different illuminations, different backgrounds, and different scale conditions; The labeled satellite image dataset contains spatial non-cooperative target images under a variety of complex and harsh environments coupled with local overexposure, feature loss, different scales, image blur and target rotation.

5. The non-cooperative target component intelligent perception method according to claim 4, characterized in that: Screening a plurality of non-cooperative target images in the space includes the following process: Performing image global contrast normalization preprocessing on several original spatial non-cooperative target images collected; Performing image scaling and data augmentation based on optimized classification on the preprocessed plurality of space non-cooperative target images to effectively expand and screen the plurality of space non-cooperative target images to obtain the labeled satellite image data set; The global contrast normalization preprocessing of the collected original several spatial non-cooperative target images includes the following process: Assume that the original spatial non-cooperative target image is represented by a tensor: Where X represents the intensity of the image; represents the tensor representation symbol; m represents the image row; n represents the image column; The contrast of the entire image is expressed as follows: In the formula, the value range of k is 1 to 3. When k = 1, X i,j,1 represents the intensity of red in the i-th row and j-th column; when k = 2, X i,j,2 represents the intensity of green in row i and column j; when k = 3, X i,j,3 Represents the intensity of blue in row i and column j; is the average intensity of the entire image, satisfying: Image generated by global contrast normalization It is expressed by the following formula: where ε = 10 -9 , λ = 0, s = 1; The data augmentation based on optimized classification first calculates the classification accuracy of each category in the test set, and then finds a category with the lowest accuracy in the test set.

6. The non-cooperative target component intelligent perception method according to claim 5, characterized in that: The step S21 comprises: The step of dynamically collecting images of non-cooperative target components using the robot's end hand-eye camera comprises: In a robot-assisted non-cooperative target dynamic capture test system in a ground simulation environment, the end of the robot's mechanical arm is controlled to move from far to near, and the end hand-eye camera of the robot is used to dynamically collect image data of non-cooperative target components of the current operation task during the 2m-0.2m approach process; The number of image data of each component in the non-cooperative target component assembly is 100.

7. The non-cooperative target component intelligent perception method according to claim 6, characterized in that: The step S3 comprises: Step S31, using the target rapid detection network model to rapidly train the image data stream acquired in real time by the end of the robot's mechanical arm, and obtaining the state information parameters of the non-cooperative target component in real time; Step S32, performing online annotation correction on the images with confidence levels lower than 95% in the state information parameters and re-optimizing and training the parameters of the target rapid detection network model; Step S33, adopt online incremental learning optimization based on super-participation weight optimization knowledge distillation joint loss under time constraints to train the target rapid detection network model, and update and estimate the model parameters of the sample stream data obtained in real time for the current task one by one in real time.

8. An electronic device, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 7 is implemented.

9. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images