CN114913392A - A kind of high-altitude parabolic detection method and system - Google Patents
A kind of high-altitude parabolic detection method and systemDownload PDFInfo
- Publication number
- CN114913392A CN114913392ACN202210590347.0ACN202210590347ACN114913392ACN 114913392 ACN114913392 ACN 114913392ACN 202210590347 ACN202210590347 ACN 202210590347ACN 114913392 ACN114913392 ACN 114913392A
- Authority
- CN
- China
- Prior art keywords
- model
- training
- fusion
- altitude parabolic
- task
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
Description
Translated fromChinese技术领域technical field
本发明属于图像处理技术领域，尤其涉及一种高空抛物检测方法及系统。The invention belongs to the technical field of image processing, and in particular relates to a high-altitude parabola detection method and system.
背景技术Background technique
本部分的陈述仅仅是提供了与本发明相关的背景技术信息，不必然构成在先技术。The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
传统的高空抛物检测方法需要先采集高空抛物的检测图像数据并对其进行标注，然后进行训练使用。由于传统的高空抛物检测算法需要大量的高空抛物数据集，所以在训练检测网络模型的过程中，数据量的多少影响了高空抛物模型准确率的高低。The traditional high-altitude parabola detection method needs to collect the detection image data of high-altitude parabola and label it, and then use it for training. Since the traditional high-altitude parabola detection algorithm requires a large number of high-altitude parabola data sets, in the process of training the detection network model, the amount of data affects the accuracy of the high-altitude parabola model.
各个数据拥有方出于数据安全或者数据隐私的考虑，使得数据难以被公开，另外，高空抛物数据比较少，造成了实际可以使用的高空抛物数据比较少，阻碍了高空抛物检测算法的实现。Due to the consideration of data security or data privacy, each data owner makes it difficult to disclose the data. In addition, the high-altitude parabolic data is relatively small, resulting in less high-altitude parabolic data that can actually be used, which hinders the realization of high-altitude parabolic detection algorithms.
普通的联邦学习虽然可以在不共享数据的情况下进行模型的训练过程，但是参数服务器在采取同步方式融合模型，存在长时间等待的问题，这种问题在训练节点训练差异巨大时现场更加严重，而采用异步融合方式，又会存在节点落后的问题，即某个节点已经落后于整体的全局模型，使得直接融合的当时显然不合理。Although ordinary federated learning can train the model without sharing data, the parameter server uses a synchronous method to integrate the model, and there is a problem of long waiting. This problem is more serious when the training difference between the training nodes is huge. With the asynchronous fusion method, there will be the problem of node backwardness, that is, a certain node has already fallen behind the overall global model, making the direct fusion obviously unreasonable at the time.
相比已有的联邦学习方法中处理落后节点的方式，还可以直接采用权重衰减的方式，即越落后的模型权重越小，虽然一定程度上解决了节点落后的问题，但是显然直接采取降低权重的方式依旧存在不合理的地方，当该节点落后的程度很大时，其权重会趋向于0，从而忽略了该模型。Compared with the way of dealing with backward nodes in the existing federated learning method, the method of weight attenuation can also be directly adopted, that is, the more backward the model has, the smaller the weight. Although the problem of backward nodes is solved to a certain extent, it is obvious to directly reduce the weight. There is still an unreasonable place in the way of the node, when the node is far behind, its weight will tend to 0, thus ignoring the model.
总之，联邦学习方法在不需要共享用户数据的隐私数据的基础上完成对于高空抛物检测方法，一定程度上解决了高空抛物数据稀少的问题，但是采用同步的融合方式存在节点等待问题，即训练快的需要等待训练慢的节点，而采用异步的融合方式又存在节点落后的问题。In a word, the federated learning method completes the high-altitude parabolic detection method without sharing the private data of user data, which solves the problem of the scarcity of high-altitude parabolic data to a certain extent, but the synchronous fusion method has the problem of node waiting, that is, the training is fast It needs to wait for nodes with slow training, and the asynchronous fusion method has the problem of node backwardness.
发明内容SUMMARY OF THE INVENTION
为克服上述现有技术的不足，本发明提供了一种高空抛物检测方法，采用异步联邦学习加上梯度修正的方式，解决了节点落后的问题，同时保留了联邦学习的优点，扩充了数据集，保护了数据隐私。In order to overcome the above-mentioned shortcomings of the prior art, the present invention provides a high-altitude parabola detection method, which adopts asynchronous federated learning and gradient correction to solve the problem of backward nodes, while retaining the advantages of federated learning and expanding the data set. , which protects data privacy.
为实现上述目的，本发明的一个或多个实施例提供了如下技术方案：To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:
第一方面，公开了一种高空抛物检测方法，包括：In a first aspect, a high-altitude parabolic detection method is disclosed, including:
高空抛物数据预处理，将高空抛物图像数据进行处理生成的训练数据；High-altitude parabola data preprocessing, processing high-altitude parabola image data to generate training data;
将经过上述步骤生成的训练数据输入到网络模型中，利用网络的异步联邦学习生成指定的预测输出，包含两部分输出：坐标信息以及类别信息；Input the training data generated by the above steps into the network model, and use the asynchronous federated learning of the network to generate the specified prediction output, which includes two parts of output: coordinate information and category information;
基于上一步生成的预测输出值与真值进行损失计算，分别需要计算坐标损失以及类别信息的损失值，按照一定的加权方式进行组合，求解出梯度后反向更新模型的参数；The loss calculation is performed based on the predicted output value and the true value generated in the previous step. The coordinate loss and the loss value of the category information need to be calculated respectively, and they are combined according to a certain weighting method, and the parameters of the model are reversely updated after the gradient is solved;
重复上述的训练过程，训练检测网络直至网络模型训练收敛；Repeat the above training process to train the detection network until the network model training converges;
将待检测的高空抛物数据输入至训练后的模型，获得高空抛物的坐标信息以及类别信息。Input the high-altitude parabola data to be detected into the trained model, and obtain the coordinate information and category information of the high-altitude parabola.
进一步的技术方案，利用网络的异步联邦学习生成指定的预测输出，具体步骤：A further technical solution uses the asynchronous federated learning of the network to generate the specified prediction output. The specific steps are as follows:
联邦学习任务发起方向训练节点发起联邦学习任务；The federated learning task initiates the training node to initiate the federated learning task;
训练节点基于获取到指定的联邦学习任务在初始训练时直接加载初始化模型，在其他的轮次加载经过参数服务器采取梯度修正的融合策略之后进行生成的模型；The training node directly loads the initialization model during initial training based on the specified federated learning task, and loads the model generated after the parameter server adopts the gradient-corrected fusion strategy in other rounds;
加载模型之后开启高空抛物检测网络的训练，完成设定的训练轮次后将模型上传至参数服务器；After loading the model, start the training of the high-altitude parabolic detection network, and upload the model to the parameter server after completing the set training rounds;
参数服务器维护一个融合队列，按照请求融合的顺序进行模型的融合任务，无需等待其他节点完成训练任务；The parameter server maintains a fusion queue, and performs model fusion tasks in the order of requested fusion, without waiting for other nodes to complete the training tasks;
参数服务器将融合生成的模型进行下发，传递至发起该请求任务的训练节点；The parameter server delivers the model generated by fusion to the training node that initiated the request task;
该训练节点接受其下发的模型，继续进行训练任务，直至最后模型完成收敛，利用该模型完成高空抛物的检测。The training node accepts the model issued by it, and continues the training task until the final model completes the convergence, and uses the model to complete the detection of high-altitude parabolas.
本发明采用异步联邦学习加上梯度修正的方式，解决了节点落后的问题，同时保留了联邦学习的优点，扩充了数据集，保护了数据隐私。The invention adopts the method of asynchronous federated learning and gradient correction to solve the problem of node backwardness, while retaining the advantages of federated learning, expanding the data set, and protecting data privacy.
进一步的技术方案，训练节点获取到指定的联邦学习任务后，初始化联邦学习环境；之后，训练节点对于高空抛物检测的图像数据进行数据预处理。In a further technical solution, after the training node obtains the specified federated learning task, the federated learning environment is initialized; after that, the training node performs data preprocessing on the image data of high-altitude parabolic detection.
进一步的技术方案，初始化联邦学习环境，包括：初始化网络架构模型、融合模型方法、融合迭代次数、学习率及优化器参数。A further technical solution is to initialize the federated learning environment, including: initializing the network architecture model, fusion model method, fusion iteration times, learning rate and optimizer parameters.
进一步的技术方案，训练节点对于高空抛物检测的图像数据进行数据预处理，包括数据增强及图像归一化。In a further technical solution, the training node performs data preprocessing on the image data of high-altitude parabolic detection, including data enhancement and image normalization.
进一步的技术方案，所述参数服务器采用梯度修正方式进行模型融合。In a further technical solution, the parameter server adopts a gradient correction method to perform model fusion.
进一步的技术方案，采用梯度修正方式进行模型融合，将每个模型的梯度进行修正，从而促使每一个落后的模型进行梯度修正。In a further technical solution, a gradient correction method is used to perform model fusion, and the gradient of each model is corrected, thereby prompting each backward model to perform gradient correction.
进一步的技术方案，将每个模型的梯度进行修正，具体为：A further technical solution is to correct the gradient of each model, specifically:
一训练节点上传某一轮的模型；A training node uploads the model of a certain round;
计算出其梯度；Calculate its gradient;
在计算出其修正的梯度后，采用泰勒公式，展开到第二项，后面的项全部忽略；After calculating its corrected gradient, use Taylor's formula to expand to the second item, and all the following items are ignored;
采用上述修正的梯度进行模型的更新操作，生成融合后的模型。The above-mentioned modified gradient is used to update the model to generate a fused model.
第二方面，公开了一种高空抛物检测系统，包括：In a second aspect, a high-altitude parabolic detection system is disclosed, including:
高空抛物数据预处理模块，将高空抛物图像数据进行处理生成的训练数据；The high-altitude parabola data preprocessing module is the training data generated by processing the high-altitude parabola image data;
模型训练模块，将经过上述步骤生成的训练数据输入到网络模型中，利用网络的异步联邦学习生成指定的预测输出，包含两部分输出：坐标信息以及类别信息；The model training module inputs the training data generated through the above steps into the network model, and uses the asynchronous federated learning of the network to generate the specified prediction output, including two parts of output: coordinate information and category information;
基于上一步生成的预测输出值与真值进行损失计算，分别需要计算坐标损失以及类别信息的损失值，按照一定的加权方式进行组合，求解出梯度后反向更新模型的参数；The loss calculation is performed based on the predicted output value and the true value generated in the previous step. The coordinate loss and the loss value of the category information need to be calculated respectively, and they are combined according to a certain weighting method, and the parameters of the model are reversely updated after the gradient is solved;
重复上述的训练过程，训练检测网络直至网络模型训练收敛；Repeat the above training process to train the detection network until the network model training converges;
检测模块，将待检测的高空抛物数据输入至训练后的模型，获得高空抛物的坐标信息以及类别信息。The detection module inputs the high-altitude parabola data to be detected into the trained model, and obtains coordinate information and category information of the high-altitude parabola.
以上一个或多个技术方案存在以下有益效果：One or more of the above technical solutions have the following beneficial effects:
本发明采用异步联邦学习加上梯度修正的方式，解决了节点落后的问题，同时保留了联邦学习的优点，扩充了数据集，保护了数据隐私，保证了异步融合方式的合理性，提升了模型的泛化能力。The invention adopts the method of asynchronous federated learning and gradient correction to solve the problem of backward nodes, while retaining the advantages of federated learning, expanding the data set, protecting data privacy, ensuring the rationality of the asynchronous fusion method, and improving the model. generalization ability.
本发明采用联邦学习的训练机制使用更多的高空检测图像数据进行模型训练，提高模型的准确性。The present invention adopts the training mechanism of federated learning and uses more high-altitude detection image data for model training, thereby improving the accuracy of the model.
本发明采用异步联邦学习算法，训练完模型立即参与融合，无需等待其他节点任务完成，从而提高了模型的通信效率。The invention adopts an asynchronous federated learning algorithm, and the model participates in the fusion immediately after training, without waiting for the completion of other node tasks, thereby improving the communication efficiency of the model.
本发明附加方面的优点将在下面的描述中部分给出，部分将从下面的描述中变得明显，或通过本发明的实践了解到。Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will become apparent from the description which follows, or may be learned by practice of the invention.
附图说明Description of drawings
构成本发明的一部分的说明书附图用来提供对本发明的进一步理解，本发明的示意性实施例及其说明用于解释本发明，并不构成对本发明的不当限定。The accompanying drawings forming a part of the present invention are used to provide further understanding of the present invention, and the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, and do not constitute an improper limitation of the present invention.
图1为本发明实施例方法的流程图。FIG. 1 is a flowchart of a method according to an embodiment of the present invention.
具体实施方式Detailed ways
应该指出，以下详细说明都是示例性的，旨在对本发明提供进一步的说明。除非另有指明，本文使用的所有技术和科学术语具有与本发明所属技术领域的普通技术人员通常理解的相同含义。It should be noted that the following detailed description is exemplary and intended to provide further explanation of the invention. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
需要注意的是，这里所使用的术语仅是为了描述具体实施方式，而非意图限制根据本发明的示例性实施方式。It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present invention.
在不冲突的情况下，本发明中的实施例及实施例中的特征可以相互组合。Embodiments of the invention and features of the embodiments may be combined with each other without conflict.
实施例一Example 1
关于联邦学习：联邦学习是一种带有隐私保护、安全加密技术的分布式机器学习框架，旨在让分散的各参与方在满足不向其他参与者披露隐私数据的前提下，协作进行机器学习的模型训练。方法具体步骤如下：About Federated Learning: Federated learning is a distributed machine learning framework with privacy protection and secure encryption technology, which aims to allow decentralized participants to collaborate on machine learning without disclosing private data to other participants. model training. The specific steps of the method are as follows:
任务发布方发布联邦学习任务；The task issuer publishes the federated learning task;
训练参与方在各自环境下安装初始化的模型，每个参与方拥有相同的模型，依据每个训练参与方的数据，参与方可以使用当地的数据训练模型。训练完指定的次数后，上传模型至模型融合方。The training participants install the initialized models in their respective environments. Each participant has the same model. According to the data of each training participant, the participants can use the local data to train the model. After training the specified number of times, upload the model to the model fusion party.
模型融合方在收集到训练参与方上传的模型后开始采用一定的策略进行模型的融合，生成一个全局的模型，然后发送该全局模型给训练参与方。After collecting the models uploaded by the training participants, the model fusion party uses a certain strategy to fuse the models, generates a global model, and then sends the global model to the training participants.
训练参与方接收模型继续模型的训练任务，重复上述过程直至模型收敛。The training participant receives the model and continues the training task of the model, and repeats the above process until the model converges.
本发明的高空抛物检测方法基于目标检测方法，当出现高空抛物时需要检测出抛物的位置与类别。其具体实现步骤如下：The high-altitude parabola detection method of the present invention is based on the target detection method, and when a high-altitude parabola appears, the position and category of the parabola need to be detected. The specific implementation steps are as follows:
高空抛物数据预处理，将高空抛物图像数据进行处理，如进行数据增强，数据归一化等操作。High-altitude parabolic data preprocessing, processing high-altitude parabolic image data, such as data enhancement, data normalization and other operations.
将经过上述步骤生成的训练数据输入到网络中，如MaskRCNN算法中，生成指定的预测输出，包含两部分输出，坐标信息以及类别信息。Input the training data generated by the above steps into the network, such as the MaskRCNN algorithm, to generate the specified prediction output, including two parts of output, coordinate information and category information.
基于上一步生成的预测输出与真值进行损失计算，分别需要计算坐标损失以及类别信息的损失值，按照一定的加权方式进行组合，并采用BP求解出梯度后反向更新模型的参数。To calculate the loss based on the predicted output and the true value generated in the previous step, it is necessary to calculate the coordinate loss and the loss value of the category information, respectively, and combine them according to a certain weighting method, and use BP to solve the gradient and reversely update the parameters of the model.
重复上述的过程，训练检测网络直至网络训练收敛。Repeat the above process to train the detection network until the network training converges.
高空抛物的常规算法方法如上所示，本实施例子的高空抛物检测网络的异步联邦学习方法表示用联邦学习的方式训练高空抛物模型。The conventional algorithm method of the high-altitude parabola is as shown above, and the asynchronous federated learning method of the high-altitude parabola detection network of this embodiment means that the high-altitude parabola model is trained by means of federated learning.
具体的实施例子中，高空抛物检测网络的异步联邦学习方法，参见附图1所示：In a specific implementation example, the asynchronous federated learning method of the high-altitude parabolic detection network is shown in Figure 1:
(1)、当联邦学习任务发起方发起联邦学习任务，由训练节点获取到指定的联邦学习任务后，初始化联邦学习环境，通常初始化网络架构模型、融合模型方法，以及融合迭代次数，学习率，优化器等参数。、发起方表示这个高空抛物算法训练任务的创建者。训练节点通常表示这个高空抛物的数据拥有方，他使用自己的本地数据训练模型。参数服务器表示用于接收所有训练节点传输的模型参数，并将其进行融合。(1) When the initiator of the federated learning task initiates the federated learning task, after the specified federated learning task is acquired by the training node, the federated learning environment is initialized, usually the network architecture model, the fusion model method, the number of fusion iterations, the learning rate, optimizer and other parameters. , The initiator represents the creator of this high-altitude parabolic algorithm training task. The training node usually represents the data owner of this high-altitude parabola, who uses his own local data to train the model. The parameter server represents the model parameters that are used to receive transmissions from all training nodes and fuse them.
(2)、随后训练节点对于高空抛物检测的图像数据进行数据预处理，根据算法需要进行一定的训练前准备，如进行数据增强、图像归一化等。(2) Then the training node performs data preprocessing on the image data of high-altitude parabolic detection, and performs certain pre-training preparations according to the needs of the algorithm, such as data enhancement, image normalization, etc.
(3)、在初始训练时直接加载初始化模型，而在其他的轮次加载经过参数服务器采取梯度修正的融合策略之后进行生成的模型。(3) The initialization model is directly loaded during the initial training, and the generated model is loaded after the parameter server adopts the gradient-corrected fusion strategy in other rounds.
模型的参数初始化，即从数据分布中，例如高斯分布中随机采样填充相应的值。The parameters of the model are initialized, i.e. randomly sampled and filled with corresponding values from a data distribution, such as a Gaussian distribution.
(4)、随后开启高空抛物检测网络的训练过程，完成一定的训练轮次后进行模型(梯度)的上传任务。(4) Then start the training process of the high-altitude parabolic detection network, and perform the uploading task of the model (gradient) after completing a certain training round.
(5)、参数服务器维护一个融合队列，按照请求融合的顺序进行模型的融合任务，无需等待其他节点完成训练任务，并采用梯度修正方式进行模型融合。每个训练节点训练完一次之后会上传模型，上传的模型会按照队列的方式进行组织。(5) The parameter server maintains a fusion queue, performs model fusion tasks in the order of requested fusion, without waiting for other nodes to complete the training tasks, and uses gradient correction to perform model fusion. After each training node is trained once, the model will be uploaded, and the uploaded model will be organized in a queue.
采用加权融合的方式：即获取前文模型队列中的模型参数，按照一定的权重加到全局模型参数上。The method of weighted fusion is adopted: that is, the model parameters in the previous model queue are obtained and added to the global model parameters according to a certain weight.
ws＝(1-λ)*ws+λ*wcws =(1-λ)*ws +λ*wc
ws表示全局的服务器的模型参数，wc表示从队列中获取的客户端模型参数。λ表示融合权重。ws represents the global server model parameters, and wc represents the client model parameters obtained from the queue. λ represents the fusion weight.
(6)、参数服务器融合生成的模型进行下发，传递给发起该请求任务的节点。(6) The model generated by the fusion of the parameter server is delivered to the node that initiates the request task.
(7)、该训练节点接受其下发的模型，按照步骤3-7继续进行训练任务，直至最后模型完成收敛。(7) The training node accepts the model issued by it, and continues to perform the training task according to steps 3-7 until the model finally converges.
梯度修正融合方法：Gradient correction fusion method:
梯度修正的融合方法，即解决训练节点上传的模型或者梯度与参数服务器不对应的问题，也即当第t轮的参数服务器全局模型为wt，而训练节点K上传的第tk轮的模型
首先计算出其梯度：First calculate its gradient:
在计算出其修正的梯度，采用泰勒公式，展开到第二项，为了考虑计算量的问题，直接展开到第二项，后面的项全部忽略。After calculating its corrected gradient, use Taylor's formula to expand to the second item. In order to consider the problem of calculation amount, directly expand to the second item, and ignore all the following items.
采用上述修正的梯度进行模型的更新操作,生成融合后的模型。The above-mentioned modified gradient is used to update the model to generate a fused model.
wt+1＝wt-ηt*gt(wt)wt+1 =wt -ηt *gt (wt )
本发明采用联邦学习整合高空抛物检测算法，采用异步联邦学习在训练过程中传递模型，而不需要共享用户的数据，从而保护高空抛物检测的数据安全。The present invention adopts federated learning to integrate the high-altitude parabolic detection algorithm, and adopts asynchronous federated learning to transfer the model in the training process without sharing user data, thereby protecting the data security of high-altitude parabolic detection.
由于上述对于高空抛物数据安全性的保护，使得用户可以安全参与到模型的训练过程中，进一步扩充了高空抛物的数据集，增加了高空抛物的数据，从而提高了模型的准确率与泛化能力。Due to the above-mentioned protection of high-altitude parabolic data security, users can safely participate in the training process of the model, further expanding the high-altitude parabolic data set and increasing the high-altitude parabolic data, thereby improving the accuracy and generalization ability of the model. .
本发明采用异步的联邦融合方式，使得每个节点无需等待就可以完成一轮融合，从而解决了节点等待的问题，从而提升了每个训练参与方的通行效率。由于采用梯度修正的方式，将每个模型的梯度进行修正，从而促使每一个落后的模型进行梯度修正，使得异步联邦学习的合理性。The invention adopts an asynchronous federation fusion method, so that each node can complete a round of fusion without waiting, thereby solving the problem of node waiting, thereby improving the traffic efficiency of each training participant. Due to the gradient correction method, the gradient of each model is corrected, thereby prompting each backward model to perform gradient correction, making asynchronous federated learning reasonable.
另一实施例子中，采用群体学习的方法去替代联邦学习方法，即在联邦学习中存在一个融合方去不断的融合每个训练参与方的模型，群体学习基于去中心化的方式，模型的聚合可以发生在每个训练参与方上。In another embodiment, the method of group learning is used to replace the method of federated learning, that is, in the federated learning, there is a fusion party to continuously integrate the models of each training participant. The group learning is based on a decentralized method, and the aggregation of models Can happen to every training participant.
实施例二Embodiment 2
本实施例的目的是提供一种计算装置,包括存储器、处理器及存储在存储器上并可在处理器上运行的计算机程序，所述处理器执行所述程序时实现上述方法的步骤。The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and running on the processor, and the processor implements the steps of the above method when the processor executes the program.
实施例三Embodiment 3
本实施例的目的是提供一种计算机可读存储介质。The purpose of this embodiment is to provide a computer-readable storage medium.
一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时执行上述方法的步骤。A computer-readable storage medium having a computer program stored thereon, the program executing the steps of the above method when executed by a processor.
实施例四Embodiment 4
本实施例的目的是提供一种高空抛物检测系统，包括：The purpose of this embodiment is to provide a high-altitude parabolic detection system, including:
高空抛物数据预处理模块，将高空抛物图像数据进行处理生成的训练数据；The high-altitude parabola data preprocessing module is the training data generated by processing the high-altitude parabola image data;
模型训练模块，将经过上述步骤生成的训练数据输入到网络模型中，利用网络的异步联邦学习生成指定的预测输出，包含两部分输出：坐标信息以及类别信息；The model training module inputs the training data generated through the above steps into the network model, and uses the asynchronous federated learning of the network to generate the specified prediction output, including two parts of output: coordinate information and category information;
基于上一步生成的预测输出值与真值进行损失计算，分别需要计算坐标损失以及类别信息的损失值，按照一定的加权方式进行组合，求解出梯度后反向更新模型的参数；The loss calculation is performed based on the predicted output value and the true value generated in the previous step. The coordinate loss and the loss value of the category information need to be calculated respectively, and they are combined according to a certain weighting method, and the parameters of the model are reversely updated after the gradient is solved;
重复上述的训练过程，训练检测网络直至网络模型训练收敛；Repeat the above training process to train the detection network until the network model training converges;
检测模块，将待检测的高空抛物数据输入至训练后的模型，获得高空抛物的坐标信息以及类别信息。The detection module inputs the high-altitude parabola data to be detected into the trained model, and obtains coordinate information and category information of the high-altitude parabola.
在一实施例子中，模型训练模块中利用网络的异步联邦学习生成指定的预测输出，包括：In an embodiment, the model training module utilizes the asynchronous federated learning of the network to generate the specified prediction output, including:
初始化模块，被配置为：联邦学习任务发起方向训练节点发起联邦学习任务；The initialization module is configured as: the federated learning task initiates the training node to initiate the federated learning task;
训练节点基于获取到指定的联邦学习任务在初始训练时直接加载初始化模型，在其他的轮次加载经过参数服务器采取梯度修正的融合策略之后进行生成的模型；The training node directly loads the initialization model during initial training based on the specified federated learning task, and loads the model generated after the parameter server adopts the gradient-corrected fusion strategy in other rounds;
融合训练模块，被配置为：加载模型之后开启高空抛物检测网络的训练，完成设定的训练轮次后将模型上传至参数服务器；The fusion training module is configured to: start the training of the high-altitude parabolic detection network after loading the model, and upload the model to the parameter server after completing the set training rounds;
参数服务器维护一个融合队列，按照请求融合的顺序进行模型的融合任务，无需等待其他节点完成训练任务；The parameter server maintains a fusion queue, and performs model fusion tasks in the order of requested fusion, without waiting for other nodes to complete the training tasks;
参数服务器将融合生成的模型进行下发，传递至发起该请求任务的训练节点；The parameter server delivers the model generated by fusion to the training node that initiated the request task;
该训练节点接受其下发的模型，继续进行训练任务，直至最后模型完成收敛，利用该模型完成高空抛物的检测。The training node accepts the model issued by it, and continues the training task until the final model completes the convergence, and uses the model to complete the detection of high-altitude parabolas.
以上实施例二、三和四的装置中涉及的各步骤与方法实施例一相对应，具体实施方式可参见实施例一的相关说明部分。术语“计算机可读存储介质”应该理解为包括一个或多个指令集的单个介质或多个介质；还应当被理解为包括任何介质，所述任何介质能够存储、编码或承载用于由处理器执行的指令集并使处理器执行本发明中的任一方法。The steps involved in the apparatuses of the second, third, and fourth embodiments above correspond to the method embodiment 1, and the specific implementation can refer to the relevant description part of the embodiment 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more sets of instructions; it should also be understood to include any medium capable of storing, encoding or carrying for use by a processor The executed instruction set causes the processor to perform any of the methods of the present invention.
本领域技术人员应该明白，上述本发明的各模块或各步骤可以用通用的计算机装置来实现，可选地，它们可以用计算装置可执行的程序代码来实现，从而，可以将它们存储在存储装置中由计算装置来执行，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。本发明不限制于任何特定的硬件和软件的结合。Those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by the computing device, so that they can be stored in a storage device. The device is executed by a computing device, or they are separately fabricated into individual integrated circuit modules, or multiple modules or steps in them are fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.
上述虽然结合附图对本发明的具体实施方式进行了描述，但并非对本发明保护范围的限制，所属领域技术人员应该明白，在本发明的技术方案的基础上，本领域技术人员不需要付出创造性劳动即可做出的各种修改或变形仍在本发明的保护范围以内。Although the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, they do not limit the scope of protection of the present invention. Those skilled in the art should understand that on the basis of the technical solutions of the present invention, those skilled in the art do not need to pay creative work. Various modifications or deformations that can be made are still within the protection scope of the present invention.
Claims (10)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210590347.0ACN114913392A (en) | 2022-05-27 | 2022-05-27 | A kind of high-altitude parabolic detection method and system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210590347.0ACN114913392A (en) | 2022-05-27 | 2022-05-27 | A kind of high-altitude parabolic detection method and system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114913392Atrue CN114913392A (en) | 2022-08-16 |
Family
ID=82767856
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210590347.0APendingCN114913392A (en) | 2022-05-27 | 2022-05-27 | A kind of high-altitude parabolic detection method and system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114913392A (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111931719A (en)* | 2020-09-22 | 2020-11-13 | 苏州科达科技股份有限公司 | High-altitude parabolic detection method and device |
| CN112966832A (en)* | 2021-03-31 | 2021-06-15 | 上海嗨普智能信息科技股份有限公司 | Multi-server-based federal learning system |
| CN113393495A (en)* | 2021-06-21 | 2021-09-14 | 暨南大学 | High-altitude parabolic track identification method based on reinforcement learning |
| CN113516042A (en)* | 2021-05-17 | 2021-10-19 | 江苏奥易克斯汽车电子科技股份有限公司 | High-altitude parabolic detection method, device and equipment |
| CN114332163A (en)* | 2021-12-29 | 2022-04-12 | 武汉大学 | A method and system for high-altitude parabolic detection based on semantic segmentation |
- 2022
- 2022-05-27CNCN202210590347.0Apatent/CN114913392A/enactivePending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111931719A (en)* | 2020-09-22 | 2020-11-13 | 苏州科达科技股份有限公司 | High-altitude parabolic detection method and device |
| CN112966832A (en)* | 2021-03-31 | 2021-06-15 | 上海嗨普智能信息科技股份有限公司 | Multi-server-based federal learning system |
| CN113516042A (en)* | 2021-05-17 | 2021-10-19 | 江苏奥易克斯汽车电子科技股份有限公司 | High-altitude parabolic detection method, device and equipment |
| CN113393495A (en)* | 2021-06-21 | 2021-09-14 | 暨南大学 | High-altitude parabolic track identification method based on reinforcement learning |
| CN114332163A (en)* | 2021-12-29 | 2022-04-12 | 武汉大学 | A method and system for high-altitude parabolic detection based on semantic segmentation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110427433B (en) | Block chain consensus method and storage medium | |
| CN106796533B (en) | System and method for adaptively selecting execution mode | |
| CN111723947A (en) | A training method and device for a federated learning model | |
| CN112799708A (en) | Method and system for jointly updating business model | |
| CN112926747B (en) | Method and device for optimizing business model | |
| WO2020147142A1 (en) | Deep learning model training method and system | |
| CN116643867A (en) | A Task Scheduling Method Based on Improved Chimpanzee Optimization Algorithm | |
| CN111800274B (en) | Verifiable calculation energy consumption optimization method based on block chain | |
| CN118211268A (en) | Heterogeneous federal learning privacy protection method and system based on diffusion model | |
| CN116628504A (en) | A trusted model training method based on federated learning | |
| CN111709535A (en) | Federated learning method, device, central server and medium based on evolutionary computing | |
| CN109657794A (en) | A kind of distributed deep neural network performance modelling method of queue based on instruction | |
| CN113191530B (en) | Block link point reliability prediction method and system with privacy protection function | |
| CN111597276A (en) | Entity alignment method, device and equipment | |
| CN115665869A (en) | Multi-user collaboration platform and method based on edge computing and directed acyclic graph | |
| CN117350357A (en) | Weak supervision federal learning training method, device, equipment and storage medium | |
| CN114116705A (en) | Method and device for determining contribution value of participants in joint learning | |
| CN109697511B (en) | Data reasoning method and device and computer equipment | |
| CN118171718A (en) | A federated learning method based on double-layer gradient descent | |
| CN106407932A (en) | Handwritten number recognition method based on fractional calculus and generalized inverse neural network | |
| CN114417999B (en) | A person re-identification method based on federated split learning | |
| CN118228841B (en) | Personalized federal learning training method, system and equipment based on consistency modeling | |
| CN115577794A (en) | A multi-agent reinforcement learning method, device, equipment and storage medium | |
| CN114913392A (en) | A kind of high-altitude parabolic detection method and system | |
| CN114662148A (en) | Multi-party combined training method and device for protecting privacy |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |