An unmanned system visual positioning method based on bidirectional hierarchical momentum contrast learning

Abstract

The application discloses an unmanned system visual positioning method based on bidirectional layered momentum contrast learning, and specific steps include: firstly, two images of different perspectives are encoded through a neural network model main body; secondly, a bidirectional momentum encoder and a queue are respectively constructed to encode and store picture samples; thirdly, the momentum parameter is used to update the encoder and the first-in-first-out principle is used to update the queue; finally, an InfoNCE loss function is used for back propagation to reduce the distance between the image and the positive sample and increase the distance between the image and the negative sample, so that the retrieval performance of the neural network is improved. The bidirectional momentum contrast learning method provided by the application can provide a large number of negative samples for the neural network training process, and also solves the deviation possibly caused by taking only a certain perspective as a reference during reasoning.

Description

Technical Field

[0001] This invention relates to a visual localization method for unmanned systems, and more particularly to a visual localization method for unmanned systems based on bidirectional hierarchical momentum contrastive learning. Background Technology

[0002] In cross-view image retrieval tasks, a trained neural network is often used to extract image features. Then, the similarity of features from images at different viewpoints (Ground-view and Aerial-view) is calculated, and the image pair with the highest similarity is considered a match. Therefore, this requires the neural network's feature extractor to learn the distinction between matching and non-matching image pairs during training. This process necessitates providing the query image with positive samples and a sufficient number of negative samples during training so that the neural network's feature extractor can learn discriminative features.

[0003] To date, significant progress has been made in cross-view image retrieval. However, the following problems still exist:

[0004] 1) Most methods use a triplet loss function, which results in a lower loss function value for images with large differences, which is not conducive to neural network learning.

[0005] 2) Because batch processing is used during model training, the corresponding images of other images in a small batch are usually treated as mismatched images of the current image, i.e., as negative samples of the current image. This method may result in a small number of negative samples, making it difficult for the neural network to learn discriminative features.

[0006] 3) In contrastive learning methods, negative samples are typically stored in a queue, but this is usually unidirectional. For dual-network models, this can lead to bias in the retrieval direction during validation, thus affecting the model's performance. Summary of the Invention

[0007] Purpose of the invention: The technical problem to be solved by the present invention is to provide a visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning, which addresses the shortcomings of the existing technology.

[0008] To address the aforementioned technical problems, this invention discloses a visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning, comprising the following steps:

[0009] Step 1: Preprocess the images from the two perspectives and use them as inputs to the respective networks in the neural network model; the two perspective images are the first perspective image and the second perspective image, respectively.

[0010] Step 2: Design a neural network model to extract image features and obtain image features at different stages, namely local features and global features; train the neural network model using a metric learning training method.

[0011] Step 3: Design an encoder to encode image features and create two queues to store the image features processed by the encoder from two different viewpoints.

[0012] Step 4: Treat the first-view image as an anchor point as the input to the neural network model, and the second-view image as the input to the encoder. Process the image features output by the encoder, store them in the corresponding queue, and update the queue.

[0013] Step 5: Swap the first-view and second-view images as input positions, and repeat step 4;

[0014] Step 6: Calculate the similarity between the Anchor and the positive and negative samples by using the image features of different stages output by the neural network model and the corresponding image features output by the encoder.

[0015] Step 7: Using the obtained similarity, calculate the loss function values ​​of features at different stages, perform backpropagation, and update the neural network model;

[0016] Step 8: Update encoder parameters with momentum, and repeat steps 4 to 7 above to complete encoder update and neural network model training;

[0017] Step 9: Using the trained neural network model and the updated encoder, perform visual localization to complete the visual localization of the unmanned system based on bidirectional hierarchical momentum contrast learning.

[0018] Furthermore, the first-view image and the second-view image mentioned in step 1 are respectively the ground-view and the aerial-view.

[0019] Furthermore, the preprocessing described in step 1 includes: performing polar coordinate transformation on the images from both perspectives, marking geographical locations, cropping, scaling, and data augmentation.

[0020] Furthermore, the neural network model described in step 2 includes:

[0021] The VGG16 consists of a convolutional layer, a local encoder, and a global encoder. Both the local encoder and the global encoder are composed of four Deit blocks.

[0022] Furthermore, the training described in step 2 uses InfoNCE as the loss function, as detailed below:

[0023] A temperature coefficient τ is set to control the discrimination of the neural network model against negative samples. Let the image feature of the Anchor be q, and the image from the i-th different viewpoint, i.e., the sample, be k. i Among them, the number of positive samples that match the Anchor is k. + Then the loss function L is:

[0024]

[0025] Where k represents the sample size, exp() represents the exponential function with base e as the natural logarithm, and S () This indicates similarity calculation.

[0026] Furthermore, the encoder design described in step 3 specifically includes:

[0027] The neural network model designed in step 2 is used as the encoder. This encoder does not perform backpropagation in subsequent processes, but only updates the momentum parameters.

[0028] Furthermore, step 4, which involves processing the image features output by the encoder, storing them in the corresponding queue, and updating the queue, specifically includes:

[0029] Suppose that the local features output by the encoder after processing the first-view and second-view images are respectively and The global features are respectively and After connecting by channel, the cascaded features of the first-view and second-view images are as follows: and Specifically, it is expressed as follows:

[0030]

[0031]

[0032] The above cascaded features and Store them into the corresponding queues, and update the queues according to the first-in, first-out principle.

[0033] Furthermore, the similarity calculation described in step 6 specifically includes:

[0034] S (a,g) =f a ·f g

[0035] Among them, S (a,g) f represents the similarity. a and f gFor different image features, '·' represents the dot product operation between vectors.

[0036] Furthermore, the loss function value L mentioned in step 7 total The calculation method is as follows:

[0037] L total =α h L h +α g L g +α c L c

[0038] Wherein, the local feature loss is L h Its loss weight is α h The global feature loss is L. g Its loss weight is α g The cascaded feature loss is L c Its loss weight is α c ;

[0039] The loss function values ​​for the features at different stages are obtained using the similarity calculation method described in step 6 and the loss function calculation method described in step 2.

[0040] Furthermore, the momentum update encoder parameters described in step 8 specifically include:

[0041] Let the neural network for the first-person view image be G. q Its encoder is G k The second-view image neural network is A. q Its encoder is A k The encoder parameter is θ, and the momentum update process is as follows:

[0042]

[0043]

[0044] The hyperparameter m∈[0,1) is set to control the speed of the encoder momentum update.

[0045] Beneficial effects:

[0046] 1. The present invention provides a bidirectional hierarchical momentum contrast learning method for visual localization of unmanned systems. By comprehensively considering the differences between cross-view images and the hierarchical neural network model, it can achieve better results than general training methods and realize image matching from different viewpoints.

[0047] 2. The algorithm provided by this invention uses a more discriminative training method to obtain sufficiently distinguishable features.

[0048] 3. This invention also employs a bidirectional hierarchical contrastive learning method, which can be used for more complex models. It also compensates for the shortcomings of commonly used retrieval methods, which may lead to information flow bias. Attached Figure Description

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0050] Figure 1 This is a schematic diagram of the overall process in this invention.

[0051] Figure 2 This is a schematic diagram of the network structure of the cross-view training part in this invention.

[0052] Figure 3 This is a schematic diagram of the network structure for extracting image features according to the present invention.

[0053] Figure 4 This is a schematic diagram of the comparison learning queue access process in this invention.

[0054] Figure 5 This is a complete schematic diagram of the loss function calculated using a queue in this invention. Detailed Implementation

[0055] This invention provides a bidirectional hierarchical momentum contrastive learning method for visual localization in unmanned systems. This method increases the number of negative samples, making the number of negative samples independent of batch size. Furthermore, the bidirectional contrastive learning approach ensures that network models from different perspectives can obtain sufficient negative samples for training, reducing bias in retrieval direction during validation and simultaneously improving the network's retrieval performance.

[0056] The technical solution of this invention is: a bidirectional hierarchical momentum contrast learning method for visual localization in unmanned systems, specifically including the following steps:

[0057] Step 1: During the training phase, images from both Ground-view and Aerial-view perspectives are preprocessed and used as inputs to their respective networks;

[0058] Step 2: Design a neural network model to process images from different perspectives, extracting image features and obtaining local and global feature outputs. A metric learning training method is used to enable the network to learn more discriminative features. InfoNCE is used as the loss function, and a temperature coefficient τ is set to control the network model's discriminative power against negative samples. Let the image features of the Anchor be q, and the samples from different perspectives be k. i Among them, the number of positive samples that match the Anchor is k.+ Then the definition of Loss is:

[0059]

[0060] Where k represents the sample size, exp() represents the exponential function with base e as the natural logarithm, and S () This indicates similarity calculation.

[0061] Step 3: Design encoders for different viewpoints to encode image features, and create a queue for each to store the encoded image features;

[0062] Step 4: First, treat the ground-view image as the anchor and use it as the input to the corresponding neural network model. Then, use the aerial-view image as the input to the corresponding encoder. Concatenate the image features from different stages of the encoder by channel and store them in the corresponding queues. Update the image features stored in the queues according to the first-in, first-out principle. Let the local output features of the encoder from different perspectives be... The global output features are respectively Based on the cascading characteristics after channel connection, they are respectively The storage process of a queue can be represented as:

[0063]

[0064]

[0065] Step 5: Then, treat the Aerial-view image as the Anchor and use it as the input to the corresponding neural network model, and the Ground-view image as the input to the corresponding encoder, and repeat step 4.

[0066] Step 6: Calculate the similarity between the Anchor and the positive / negative samples by using image features from different stages of the neural network and feature encodings from samples at different viewpoints corresponding to those stages. Let · represent the vector dot product operation, and let f represent the features from different viewpoints. a f g The similarity between them is defined as

[0067] S (a,g) =f a ·f g

[0068] Among them, S (a,g) f represents the similarity. a and f g For different image features, '·' represents the dot product operation between vectors.

[0069] Step 7: Calculate the loss function values ​​for features at different stages using the obtained similarity scores to achieve backpropagation. Let L be the loss for partial-level features. h Its loss weight is α h The global feature loss is L. g Its loss weight is α g ;

[0070] Cascaded feature loss is L c Its loss weight is α c The overall loss function can be defined as:

[0071] L total =α h L h +α g L g +α c L c

[0072] Step 8: Update encoder parameters using momentum. Let the ground-view neural network model be G. q Its encoder is G k Aerial-view neural network is used for model A q Its encoder is A k The encoder parameter is θ, and the momentum update rate is m∈[0,1). The encoder momentum update can be expressed as:

[0073]

[0074]

[0075] As a further improvement of the present invention, in step 1, polar coordinate transformation is performed on the Aerial-view perspective to reduce the differences between images from different perspectives, and the images from different perspectives are cropped and scaled to 128. 512px.

[0076] As a further improvement of the present invention, in step 2, the network model consists of convolutional layers of VGG16 (reference: Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556,2014.), a local encoder (Part Encoder with 4 Deit-Small Blocks, Deit blocks (reference: Touvron H, Cord M, Douze M, et al. Training data-efficient image transformers & distillation through attention[C] / / International conference on machine learning.PMLR,2021:10347-10357)), and a global encoder (Global Encoder with 4 Deit-Small Blocks). Training uses InfoNCE as the loss function, which results in higher similarity between images with matching positions from different viewpoints and lower similarity between images with non-matching positions.

[0077] As a further improvement to this invention, in step 3, the neural network model designed in step 2 is used as an encoder that does not perform backpropagation; only momentum parameters are updated during training. Furthermore, a queue needs to be created for each different viewpoint to store the image features encoded by the corresponding encoder.

[0078] As a further improvement of the present invention, in step 4, the queue from step 3 is used to store image features from different stages of the encoder. To minimize the need for additional queues, image features from different stages are stored in the queues concatenated by channel.

[0079] As a further improvement of the present invention, in step 6, based on the image features from different perspectives and the encoder features from the corresponding perspectives extracted in steps 4 and 5, the image features obtained by the encoder are stored in a queue, and the queue is updated according to the first-in-first-out principle.

[0080] As a further improvement of the present invention, in step 7, the similarity and loss function values ​​are calculated respectively using the Anchor obtained in step 6 and the image features of positive and negative samples at different stages, and the backpropagation update of the training network is realized.

[0081] As a further improvement of the present invention, in step 8, the parameters of the image encoders at different viewpoints are updated using the parameters updated by the training network in step 7, thereby reducing the difference between the encoder and the training network.

[0082] Example:

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0084] Example 1: As Figure 1 The method described above, a bidirectional hierarchical momentum contrast learning method for visual localization in unmanned systems, includes the following steps:

[0085] Step 1: Data Input Stage.

[0086] Step 2: Model building phase, including Ground-view and Aerial-view models.

[0087] Step 3: Model training phase.

[0088] Step 4: Encoder update and queue update phase.

[0089] Step 5: Feature Extraction Stage

[0090] Step 6: Image retrieval stage.

[0091] During the data input phase, images from both ground-view and aerial-view perspectives are preprocessed. Images with the same geographical location are labeled, and all images are cropped and scaled. Finally, data augmentation is performed as needed.

[0092] In the model building phase, a neural network model was designed to process images from different perspectives, extracting image features and obtaining local and global feature outputs. The network model consists of VGG16 convolutional layers, a Part Encoder (4 Deit-Small Blocks), and a Global Encoder (4 Deit-Small Blocks).

[0093] During the model training phase, InfoNCE is used as the loss function, and the temperature coefficient τ is set to control the network model's ability to distinguish negative samples. The loss function of the output features at different stages is calculated.

[0094] During the encoder update and queue update phases, momentum updates the parameters of the encoders for images from different viewpoints, while simultaneously updating the sample features stored in all queues.

[0095] In the image retrieval stage, the network is used to obtain features from different perspectives for similarity calculation, and the k most similar samples are obtained based on the top k.

[0096] The specific content of each stage is described below.

[0097] (1) During the data input stage, images from both Ground-view and Aerial-view perspectives were preprocessed. Due to the significant differences in perspective between the images, to reduce the discrepancies, polar coordinate transformation was performed on the Aerial-view images based on the image alignment characteristics of the dataset. Then, the image sizes were all scaled down. To enhance the model's generalization ability, the panoramic images of the Ground-view were cropped to obtain Ground-view images with a viewing angle of 180° and 90° respectively.

[0098] (2) Model building stage, such as Figure 1 As shown, this is a network primarily used for encoding ground-view and aerial-view features. It consists of VGG16 convolutional layers, a Part Encoder (4 Deit-Small Blocks), and a Global Encoder (4 Deit-Small Blocks). The VGG16 convolutional layers flexibly encode images into sequential features, which then serve as input to the Part Encoder. This part, based on self-attention in local neighborhoods, sums the outputs of the VGG16 convolutional layers as input to the Global Encoder. The Global Encoder interacts with global regions to obtain features from three different stages: the Part Encoder, the Global Encoder, and the concatenation of the two. The output features of the Part Encoder and the Global Encoder are complementary and discriminative. Figure 2 As shown, the output dimensions of both the Part Encoder and Global Encoder are set to 384, resulting in a 768-dimensional output after concatenation. From this, the hierarchical loss function can be obtained, with weights α. h α g and α c The values ​​were set to 0.1, 0.1, and 5.0 respectively:

[0099] L total =α n L h +α g L g +α c L c

[0100] (3) Model training phase, such as Figure 3As shown, this invention transforms the cross-view image retrieval task into a metric learning problem. During training, the similarity between an image and its positive and negative samples is used for loss calculation. The aim is to increase the similarity between matching image pairs and reduce the similarity between non-matching image pairs, enabling the neural network to learn discriminative features. Since separate neural networks are designed to encode images from different viewpoints, both networks use the same network structure to save computational costs and memory space. Furthermore, to accelerate the training process, further reduce the number of model parameters, and decrease dependence on training equipment, parameter sharing is implemented between the two networks, reducing the number of parameters by half without sacrificing performance.

[0101] (4) Encoder update and queue update phases, such as Figure 1 As shown, after a batch of training, the network undergoes parameter updates. To reduce the discrepancy between the encoder and the training network, the encoder parameters are updated using momentum based on the updated network parameters. The hyperparameter *m* represents the speed of momentum updates. A larger *m* indicates that the momentum encoder parameters reference historical values ​​more; when *m* = 1, the momentum encoder stops updating; when *m* = 0, the parameters of the training network are directly copied as the encoder parameters. For stable training, *m* is set to a value close to 1, allowing encoders from different viewpoints to update simultaneously.

[0102] In addition to the two encoders, two queues need to be created, each containing local and global negative sample feature vectors from the ground-view and aerial-view images, respectively. To conserve the number of queues and facilitate access, the local and global feature vectors of the samples are concatenated by dimension and stored in the queues. Each negative sample queue can store M samples, where each sample has a dimension of 2C, i.e., the local feature vector has a dimension of C, and the global feature vector has a dimension of C. Before training, since the queues do not contain samples, they are initialized with random values.

[0103] like Figure 4 The diagram illustrates the queue update process. After each batch is processed, according to the first-in, first-out (FIFO) principle, the head B samples of the queue are popped, and the new B samples are added to the tail of the queue, thus updating the queue. By updating the queue, the difference between the negative sample set and the features generated by the current model is reduced, allowing for more effective comparison. Therefore, the queue storage in a batch can be represented as follows:

[0104]

[0105] In a single batch processing operation, it is necessary to calculate the similarity score of any pair of samples in the current sample set. h(i,j) S f(i,j)S c(i,j) This represents the similarity between image features at different stages, where i and j represent the positions of the Anchor and the image from another viewpoint in the current sample set, respectively. Taking the Ground-view image as the Anchor as an example, the similarity calculation is as follows:

[0106]

[0107] in, These represent the local features, global features, and channel-connected features output by the Anchor through the neural network model, respectively.

[0108] (5) Feature extraction stage, such as Figure 5 As shown, after the network training is completed, the test set images from different perspectives are preprocessed and then input into the trained network. The local features and global features are concatenated according to their dimensions to obtain a feature vector with a dimension of 768, which is used for the retrieval task.

[0109] (6) In the image retrieval stage, for any query image, its feature vector is compared with the similarity of all images from different perspectives in the database. The top-k similarity images from different perspectives are retrieved based on the top-k metric. If the correct matching image is among the top-k images, the retrieval task is considered successful; otherwise, it is considered a failure. The evaluation metric k is typically set to 1, 5, 10, or 1% to evaluate the retrieval performance of the network model. This retrieval method introduces a bias in information flow when finding matching images from other perspectives based on the query image. The proposed bidirectional contrastive learning method effectively reduces this potential bias. In practical applications, the real-time acquired ground view image (i.e., the query image) is input into the trained neural network model to obtain its feature vector. This feature vector is then compared with the feature vectors of all pre-extracted bird's-eye view images in the database. The retrieved images are obtained based on the top-1 metric. The ground location information of the input query image can then be obtained based on the geographical location information accompanying the retrieved images.

[0110] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a visual positioning method for unmanned systems based on bidirectional hierarchical momentum contrast learning, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0111] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MUU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0112] This invention provides an idea and method for visual localization of unmanned systems based on bidirectional hierarchical momentum contrast learning. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning, characterized in that, Includes the following steps: Step 1: Preprocess the images from the two perspectives and use them as inputs to the respective networks in the neural network model; the two perspective images are the first perspective image and the second perspective image, respectively. Step 2: Design a neural network model to extract image features and obtain image features at different stages, namely local features and global features; The neural network model is trained using a metric learning training method; Step 3: Design an encoder to encode image features and create two queues to store the image features processed by the encoder from two different viewpoints. Step 4: Treat the first-view image as an anchor point as the input to the neural network model, and the second-view image as the input to the encoder. Process the image features output by the encoder, store them in the corresponding queue, and update the queue. Step 5: Swap the first-view and second-view images as input positions, and repeat step 4; Step 6: Calculate the similarity between the image features output by the neural network model at different stages and the image features output by the encoder at the corresponding stages to obtain the similarity between the anchor and the positive and negative samples. Step 7: Using the obtained similarity, calculate the loss function values ​​of features at different stages, perform backpropagation, and update the neural network model; Step 8: Update encoder parameters with momentum, and repeat steps 4 to 7 above to complete encoder update and neural network model training; Step 9: Using the trained neural network model and the updated encoder, perform visual localization to complete the visual localization of the unmanned system based on bidirectional hierarchical momentum contrast learning.

2. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The first-view image and the second-view image mentioned in step 1 are respectively the ground-view and the aerial-view.

3. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The preprocessing described in step 1 includes: polar coordinate transformation of the images from both perspectives, marking geographical locations, cropping, scaling, and data augmentation.

4. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The neural network model described in step 2 includes: The VGG16 consists of a convolutional layer, a local encoder, and a global encoder. Both the local encoder and the global encoder are composed of four Deit blocks.

5. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The training described in step 2 uses InfoNCE as the loss function, as detailed below: Set temperature coefficient To control the discriminative power of the neural network model against negative samples, let the image features of the Anchor be... , No. Images from different perspectives constitute the sample. The positive samples that match the Anchor are Then the loss function for: ； in, Indicates the sample size, 𝑒𝑥𝑝 Represented by the natural logarithm An exponential function with base 0. This indicates similarity calculation.

6. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The encoder design described in step 3 specifically includes: The neural network model designed in step 2 is used as the encoder. This encoder does not perform backpropagation in subsequent processes, but only updates the momentum parameters.

7. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, Step 4, which involves processing the image features output by the encoder, storing them in the corresponding queue, and updating the queue, specifically includes: Suppose that the local features output by the encoder after processing the first-view and second-view images are respectively and The global features are respectively and After connecting by channel, the cascaded features of the first-view and second-view images are as follows: and Specifically, it is expressed as follows: ； The above cascaded features and Store them into the corresponding queues, and update the queues according to the first-in, first-out principle.

8. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, Step 6, which involves similarity calculation, specifically includes: ； in, Indicates similarity. and For different image features, This represents the dot product operation between vectors.

9. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 7, characterized in that, The loss function value mentioned in step 7 The calculation method is as follows: ； Among them, the local feature loss is Its loss weight is The global feature loss is Its loss weight is Cascaded feature loss is Its loss weight is .

10. The visual localization method for unmanned systems based on bidirectional hierarchical momentum contrast learning according to claim 1, characterized in that, The momentum update encoder parameters mentioned in step 8 specifically include: Let the first-view image neural network be... Its encoder is The second-view image neural network is Its encoder is The encoder parameters are The momentum update process is as follows: ； ； Among them, setting hyperparameters This is used to control the speed of the encoder's momentum update.

Images