A low-quality image representation model training method and system with momentum-based weighted update
Through the image representation model training method of momentum weighted update, the traditional method of attention weight imbalance and parameter update unstable in low-quality image processing is solved, and better low-quality image representation effect and model stability are achieved.
Patent Information
- Application Number
- CN202410918412.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-10
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-10
AI Technical Summary
Traditional image characterization model training methods have problems such as uneven attention weight, unstable parameter updates and batch size limitations when processing low-quality images, resulting in poor representation effect and weak generalization ability of the model on low-quality images.
The low-quality image characterization model training method is adopted with momentum weighted update. By generating anchor samples, positive samples and negative samples, a self-supervised training sample unit is constructed, the loss is calculated using chain stack temporary storage technology, and the momentum update strategy is adopted, according to different properties. The weight of the sample updates the model parameters to ensure that the model remains stable during training.
It significantly improves the learning effect and generalization ability of low-quality image characterization models, can pay more effectively to the low-quality image information, and improve the stability and performance of the model.
Smart Images

Figure CN118470467B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image representation learning, and in particular to a method and system for training a low-quality image feature representation model. Background Art
[0002] In today's digital age, image recognition and computer vision technology have become an important part of all walks of life, and its applications cover many fields such as medical diagnosis, security monitoring, and autonomous driving. However, in order for computer systems to accurately recognize and understand images, the key is to train an effective image representation model. The quality of this model directly affects the performance of the system and the effect of the application.
[0003] At present, training for high-quality images has made significant progress, but when faced with low-quality images, traditional training methods often encounter a series of challenges. Low-quality images may be caused by a variety of factors, such as image blur, noise interference, low contrast, etc. These factors make the information in the image unclear, thus bringing difficulties to model training.
[0004] When dealing with low-quality images, existing methods for training image representation models often suffer from the following three main problems:
[0005] 1. Traditional training methods usually do not assign different attention weights to samples of different properties when learning the model, which results in the trained model being unable to pay more attention to low-quality sample information, and the model's representation effect on low-quality images is often unsatisfactory.
[0006] 2. Traditional parameter update methods usually only consider the gradient information at the current moment, while ignoring the impact of the state of the parameters at previous moments on the current update. This "instant" update strategy may cause the model parameters to fluctuate greatly during the training process, making it difficult to maintain stability.
[0007] 3. Traditional training methods usually use batch updates, that is, one update requires processing a batch of data. The batch size, as a hyperparameter, directly affects the training effect, especially when processing low-quality images. However, the choice of batch size may be limited by video memory capacity or other hardware resources. Summary of the invention
[0008] The present invention provides a momentum-based weighted update low-quality image representation model training method and system to solve the problems of poor performance and weak generalization ability of existing low-quality image representation models.
[0009] In order to achieve the above object, the present invention is implemented by the following technical solutions:
[0010] In a first aspect, the present invention provides a momentum-based weighted update low-quality image representation model training method, comprising the following steps:
[0011] Step 1: Given an original data set, generate anchor samples based on each image in the original data set, define its positive sample and negative sample sets, and form the generated anchor samples, the defined positive samples and negative sample sets into self-supervised training sample units. Each supervised training sample unit forms a self-supervised training data set, which is divided into a training set and a validation set according to a predetermined ratio.
[0012] Constructing an encoder model, wherein the encoder model includes a standard encoder model, an anchor sample encoder model, a positive sample encoder model, and a negative sample encoder model;
[0013] The method for forming the self-supervised training sample unit includes:
[0014] Step 101: artificially synthesize a single low-quality image corresponding to a single image in the original data set;
[0015] Step 102: the single low-quality image is defined as an anchor sample, the corresponding original image is defined as a positive sample, and all images other than the positive sample in the original data set are defined as negative samples, and the positive sample, anchor sample and negative sample are combined into a self-supervised training sample unit;
[0016] Step 2: Input the self-supervised training sample units in the training set into the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model for training;
[0017] Step 3: Combined with the chain stack temporary storage technology, calculate the training losses of the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model, and further calculate the gradient;
[0018] Step 4: Update the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model according to the gradient, calculate the standard encoder model parameters to be updated with the updated anchor sample encoder model parameters, the positive sample encoder model parameters, the negative sample encoder model parameters, and the current standard encoder model parameters and the preset weights, and finally update the standard encoder model;
[0019] Step 5: Repeat steps 2 to 4 until the predetermined training rounds are reached, then use the validation set to verify the retrieval performance of the standard encoder model, and save the best model among all current standard encoder models as the low-quality image representation model.
[0020] Furthermore, the method further includes step 6: repeating steps 2 to 5, and completing the training of the low-quality image representation model after reaching a predetermined total number of training rounds.
[0021] Furthermore, in step 2, the training process of the anchor sample encoder model, the positive sample encoder model and the negative sample encoder model is independent training;
[0022] Among them, the above independent training is that the positive sample encoder model is used for positive sample encoding, the negative sample encoder model is used for negative sample encoding, and the anchor sample encoder model is used for anchor sample encoding.
[0023] Furthermore, the standard encoder model, anchor sample encoder model, positive sample encoder model and negative sample encoder model are all composed of a deep encoder and a binary network, and have the same network structure and the same model parameters.
[0024] Furthermore, the step 3 comprises the following steps:
[0025] Step 301: Select negative samples from the self-supervised training sample unit and add them to the constructed chain stack one by one, and determine whether the chain stack is full. If not, continue to add negative samples one by one. If it is full, calculate the contrast loss of the anchor sample, positive sample in the self-supervised training sample unit and the negative sample set in the chain stack;
[0026] The space size of the chain stack is smaller than the size of the negative sample set in the self-supervised training sample unit;
[0027] Step 302: Calculate the quantization loss and distribution loss of the anchor sample, positive sample and negative sample set in the chain stack in the self-supervised training sample unit, perform weighted average calculation in combination with the contrast loss, obtain the training loss, and obtain the gradient according to the derivation of the training loss;
[0028] Specifically, the mean square error of the deep features before quantization and the binary hash codes after quantization of the anchor samples, positive samples in the self-supervised training sample unit, and the negative sample set in the chain stack is calculated. The mean square error is the quantization loss.
[0029] Since the number of samples input into the encoder model for training is limited by the server computing resources, it cannot be set very large. However, for the effect of calculating the contrast loss to punish the model, a smaller negative sample set is difficult to meet the requirements. In the present invention, the chain stack temporary storage technology can decouple the size of the negative sample set involved in learning from the number of negative samples used for training the encoder model, greatly increasing the number of negative samples involved in model learning, and having a better learning effect on the image representation model.
[0030] Furthermore, step 4 includes the following steps:
[0031] Step 401: Based on the gradient, back-propagate the gradient in the positive sample encoder model, the negative sample encoder model and the anchor sample encoder model to update the positive sample encoder model, the negative sample encoder model and the anchor sample encoder model;
[0032] Step 402: perform weighted average calculation on the updated positive sample encoder model parameters, negative sample encoder model parameters, anchor sample encoder model parameters and current standard encoder model parameters to obtain updated standard encoder model parameters and finally update the standard encoder model.
[0033] The standard encoder model is updated weightedly based on the positive sample encoder model, negative sample encoder model, and anchor sample encoder model. During the update process, different weights can be assigned to the encoder model parameters of samples of different properties to achieve the learning goal of different attention weights. Therefore, when giving greater weight to the anchor sample (low-quality image) encoder model parameters when updating the model, the standard encoder model will pay more attention to the information from the anchor sample encoder model, allowing us to learn a better low-quality image representation model. In addition, the standard encoder model is updated with momentum, that is, the model parameters of the previous time will also participate in the model parameter update this time. Using momentum update means that each training is a slow update of the standard encoder network, which is conducive to maintaining the stability of model parameters during training.
[0034] Furthermore, after executing step 4, repeat steps 3 to 4, and clear the negative samples in the chain stack before each repetition until all negative samples in the self-supervised training sample unit are selected, and then continue to execute step 5.
[0035] Furthermore, when performing weighted average calculation, the weight of the anchor sample encoder model parameters is greater than the weight of the positive sample encoder model parameters and the negative sample encoder model parameters.
[0036] In a second aspect, the present invention provides a momentum-based weighted update low-quality image representation model training system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above-mentioned methods when executing the computer program.
[0037] Beneficial effects:
[0038] The present invention provides a momentum-based weighted update low-quality image representation model training method and system. Due to the limitation of server computing resources, the number of samples input into the encoder model for training cannot be set very large. However, in order to effectively calculate the contrast loss and punish the model, a smaller negative sample set is often difficult to achieve the desired effect. In the present invention, by setting a chain stack, the size of the negative sample set is decoupled from the number of samples input into the encoder model for training, thereby greatly increasing the number of negative samples participating in model learning and significantly improving the learning effect of the image representation model.
[0039] The standard encoder model is updated weightedly based on the positive sample encoder model, the negative sample encoder model, and the anchor sample encoder model. During the update process, different weights can be assigned to the encoder model parameters of samples of different properties to achieve differentiated attention learning goals. Therefore, when the anchor sample (low-quality image) encoder model parameters are given more weight during model update, the standard encoder model will pay more attention to the information from the anchor sample encoder model, allowing us to learn a better low-quality image representation model.
[0040] In addition, the standard encoder model adopts a momentum update strategy, that is, the previous model parameters will also participate in the update of the current model parameters. Momentum update means that each training is a progressive update of the standard encoder network, which helps to maintain the stability of model parameters during training. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 A flowchart of a low-quality image representation model training method with momentum weighted update according to an embodiment of the present invention;
[0042] Figure 2 is a schematic diagram of a positive sample in an embodiment of the present invention;
[0043] Figure 3 is a schematic diagram of an anchor sample in an embodiment of the present invention;
[0044] Figure 4 Schematic diagram of a negative sample in an embodiment of the present invention. DETAILED DESCRIPTION
[0045] The technical solution of the present invention is described clearly and completely below. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate quantity restrictions, but indicate the existence of at least one. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.
[0047] See also Figure 1 The present application embodiment provides a momentum-based weighted update low-quality image representation model training method, comprising the following steps:
[0048] S100, given data set , based on the dataset For each original image in , synthesize its approximate low-quality image. The specific steps are as follows:
[0049] 1. Given a dataset , which contains images represented as ,in, Indicates the length of the dataset, and does not repeatedly select the original image ;
[0050] (ii) Image Implement the mean blur operation, that is, under the action of one-dimensional convolution kernel, shift 7 bits in the horizontal and vertical directions respectively, and output ;
[0051] (III) Select another original image in the original dataset ( ), and generate the corresponding approximate image , and so on, until there are no more images in the dataset. At this point, the artificially synthesized low-quality image set is .
[0052] S200, see Figures 2 to 4 , select a low-quality image synthesized in step S100 as the anchor sample, its corresponding original image as the positive sample, and all images in the original data set except the positive sample are defined as the negative sample set. Specifically, select an approximate low-quality image And defined as anchor sample, original image Defined as positive samples, other images in the original dataset except positive samples Defined as a negative sample set; finally, the positive sample, anchor sample and negative sample set are combined into a self-supervised sample unit For example, when hour, It is a positive sample, is the anchor sample, , , , , , …, is the set of negative samples. In other words, and are positive sample pairs. and , and , and , and , and , …, and They are mutually negative sample pairs.
[0053] S300, multiple self-supervised training sample units The length of the composition is The complete self-supervised dataset ,in In the self-supervised training sample units, self-supervised training sample units are divided into the training set, and the remaining The self-supervised training sample units are divided into validation sets.
[0054] S400, initialize the standard encoder model, anchor sample encoder model, positive sample encoder model, and negative sample encoder model. These models have the same network structure and model parameters, and are all composed of a deep encoder based on VGG-16 and a binary network implemented by a binary function. The standard encoder model, anchor sample encoder model, positive sample encoder model, and negative sample encoder model are independent during the training process, wherein the positive sample encoder model is used for positive sample encoding, the negative sample encoder model is used for negative sample encoding, and the anchor sample encoder model is used for anchor sample encoding.
[0055] S500: Input the self-supervised training sample units in the training set constructed in the above steps into the depth encoder in the encoder model one by one for depth encoding to obtain the depth features of the image. The specific steps are as follows:
[0056] S510, self-supervised training sample units ( ) , , One of for A subset of images ( and ) are input into the preprocessor in sequence for preprocessing. The specific steps are as follows:
[0057] (I) All images perform image read operations, that is, open the image handle and generate Object ;
[0058] (ii) The object performs image transformation operations, which include: transforming image channel specifications, normalizing, reconstructing data distribution, adjusting The dimensions of the object, where:
[0059] ① Change channel specifications: Change the specifications of (height, width, channel) The data type is converted to (channel, height, width) Data type;
[0060] ② Normalization: Divide the values of each dimension of height, width and channel by 255 to normalize the values to [0,1];
[0061] ③ Reconstruct data distribution: Under the condition that the input mean is [0.485, 0.456, 0.406] and the standard deviation is [0.229, 0.224, 0.225], reconstruct the numerical distribution to the interval (-1, 1);
[0062] ④ Adjust image size: Adjust The object is of standard 224×224 size;
[0063] (III) After the above preprocessing operations, high-dimensional feature data of images in the supervised training sample unit can be abstracted .
[0064] S520, deep encoding the high-dimensional feature data of the image to obtain image depth features:
[0065] ;
[0066] ;
[0067] ;
[0068] in, is the positive sample deep encoder, is the anchor sample deep encoder, is the negative sample deep encoder.
[0069] S600: Sequence the depth features Input to a nonlinear binary function, output a binary hash code The nonlinear binary function is expressed as:
[0070] ;
[0071] in, express function, which transforms real-valued features Mapped to -1 or 1. Represents the feature vector Middle real digits; Represents a binary vector consisting of 0s and 1s Middle binary digits. Length is ,therefore .
[0072] S700, calculating the training loss by using a chain stack to temporarily store negative samples in the self-supervised training sample unit, wherein the specific steps S710 to S7100 are:
[0073] S710, create a chain stack , the chain stack size is .Will The negative sample binary hash code in is added to the chain stack . Determine the chain stack Is it full? If not, Choose another for The negative sample subset is preprocessed, deeply encoded and binarized to obtain the corresponding negative sample binary hash code set, and then continuously fill it into the chain stack, repeating this step until the chain stack is full; if it is full, the negative sample binary set is represented as ( ), calculate the binary hash code of the anchor sample , positive sample binary hash code And the binary hash code of negative samples in the chain stack The contrast loss is formalized as follows:
[0074] ;
[0075] in, represents the contrast loss, represents a hyperparameter.
[0076] S720, calculate the depth feature of the positive sample before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows:
[0077] ;
[0078] in, Represents the quantization loss of positive samples;
[0079] S730, calculate the distribution loss of the positive sample after binarization:
[0080] ;
[0081] in, Represents the distribution loss of positive samples;
[0082] S740, performing weighted average calculation on the contrast loss calculated in step S710, the quantization loss calculated in step S720, and the distribution loss calculated in step S730 to obtain the positive sample encoder model loss, which is formally expressed as follows:
[0083] ;
[0084] in, , , are preset weights, and .
[0085] S750, calculate the depth feature of the anchor sample before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows:
[0086] ;
[0087] in, represents the quantization loss of the anchor sample;
[0088] S760. Calculate the distribution loss of the anchor sample after binarization:
[0089] ;
[0090] in, represents the distribution loss of the anchor sample;
[0091] S770, performing weighted average calculation on the contrast loss calculated in step S710, the quantization loss calculated in step S750, and the distribution loss calculated in step S760 to obtain the anchor sample encoder model loss, which is formally expressed as follows:
[0092] ;
[0093] in, .
[0094] S780, calculate the depth features of negative samples before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows:
[0095] ;
[0096] in, Represents the quantization loss of negative samples;
[0097] S790. Calculate the distribution loss of negative samples after binarization:
[0098] ;
[0099] in, Represents the distribution loss of negative samples;
[0100] S7100, performing weighted average calculation on the contrast loss calculated in step S710, the quantization loss calculated in step S780, and the distribution loss calculated in step S790 to obtain the negative sample encoder model loss, which is formally expressed as follows:
[0101] ;
[0102] in, .
[0103] S800, dynamically updating low-quality image representation model parameters according to training loss, wherein specific steps S810 to S840 are:
[0104] S810, after calculating the loss in step S700, start calculating the loss gradient, and back-propagate the gradient in the positive sample encoder model, the negative sample encoder model, and the anchor sample encoder model to update the network parameters, which is formally expressed as follows:
[0105] ;
[0106] ;
[0107] ;
[0108] in, represents the learning step length, , and They represent the positive sample encoder model parameters, anchor sample encoder model parameters, and negative sample encoder model parameters respectively.
[0109] S820, momentum-based weighted update of standard encoder model parameters, formally expressed as:
[0110] ;
[0111] in, , and Represents the weight parameter and its value is greater than 0 and less than 1. represents the standard encoder model parameters to be updated, Represents the current standard encoder model parameters.
[0112] S830, continue to take the next batch of negative samples from the negative sample set of the current self-supervised training sample unit, preprocess, deeply encode and binarize them, and then update the chain stack, that is, push the binary hash codes of the old batch of negative samples in the chain stack out of the stack, and push the binary hash codes of the new batch of negative samples from the stack, and repeat the above operation until the chain stack is full. After updating the chain stack, take out all negative samples in the chain stack to participate in the loss calculation and update the encoder model. Then, repeat the above operation until all negative samples in the negative sample set of the self-supervised training sample unit are taken.
[0113] S840, traverse the next self-supervised training sample unit, repeat steps S500, S600, S700 and S810, S820 and S830 in S800, until all self-supervised training sample units in the training set are traversed.
[0114] S900, repeat steps S500, S600, S700 and S800, and when the number of repetitions modulo 10 is equal to 0, verify the performance of the current standard encoder model once. The verification process is as follows: input a self-supervised training sample unit in the S300 verification set into the current standard encoder model, and learn the anchor sample binary hash code, positive sample binary hash code, and negative sample binary hash code set. Then, use the anchor sample binary hash code as the query item, and each negative sample binary hash code in the positive sample binary hash code and negative sample binary hash code set as the database search item, and calculate the cosine similarity between the query item and each search item in turn, and then obtain the cosine similarity array, that is, one search item corresponds to one cosine similarity. Then, sort the cosine similarity array from large to small according to the size of the cosine similarity, and find the index corresponding to the positive sample in the cosine similarity array, and take the reciprocal of the index as the accuracy of this search. Repeat the above verification process until all self-supervised training sample units in the S300 verification set are selected. Next, the accuracy of each search is accumulated, and then the average value is calculated. The average value can be used as the performance indicator for verifying the current standard encoder model, thus completing a complete verification. After the verification is completed, it is determined whether the current performance indicator is better than the historical optimal performance indicator. If not, no processing is performed; if so, the current standard encoder model is saved as the optimal model.
[0115] S1000, repeat step S900, when all 300 rounds of training are completed, the entire model training is completed.
[0116] An embodiment of the present application also provides a momentum-based weighted update low-quality image representation model training system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of any of the above methods when executing the computer program.
[0117] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A momentum-based weighted update method for training low-quality image representation models. It is characterized in that The steps include: Step 1: Given an original data set, generate anchor samples based on each image in the original data set, define its positive sample and negative sample sets, and form the generated anchor samples, the defined positive samples and negative sample sets into self-supervised training sample units. Each supervised training sample unit forms a self-supervised training data set, which is divided into a training set and a validation set according to a predetermined ratio. Constructing an encoder model, wherein the encoder model includes a standard encoder model, an anchor sample encoder model, a positive sample encoder model, and a negative sample encoder model; The method for forming the self-supervised training sample unit includes: Step 101: artificially synthesize a single low-quality image corresponding to a single image in the original data set; Step 102: the single low-quality image is defined as an anchor sample, the corresponding original image is defined as a positive sample, and all images other than the positive sample in the original data set are defined as negative samples, and the positive sample, anchor sample and negative sample are combined into a self-supervised training sample unit; Step 2: Input the self-supervised training sample units in the training set into the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model for training; Step 3: Combined with the chain stack temporary storage technology, calculate the training losses of the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model, and further calculate the gradient; S301. Create a chain stack , the chain stack size is ; The space size of the chain stack is smaller than the size of the negative sample set in the self-supervised training sample unit; the binary hash code set of the self-supervised training sample unit is calculated, the negative sample set in the binary hash code set is divided into a plurality of negative sample binary hash code subsets, and the hash codes in a negative sample binary hash code subset are sequentially added to the chain stack ; Judge the chain stack Is it full? If not, select another subset of negative sample binary hash codes and keep filling them into the chain stack until the chain stack is full. If it is full, calculate the contrast loss of the anchor sample binary hash code, the positive sample binary hash code, and the negative sample binary hash code in the chain stack, which is formalized as follows: ; in, represents the hyperparameter, represents the binary hash code of the anchor sample, represents the binary hash code of the positive sample, represents negative samples, represents the starting number of the negative sample set, and ; S302, calculate the depth feature of the positive sample before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows: ; S303, calculate the distribution loss of the positive sample after binarization: ; S304, performing weighted average calculation on the contrast loss calculated in step S301, the quantization loss calculated in step S302, and the distribution loss calculated in step S303 to obtain the positive sample encoder model loss, which is formally expressed as follows: ; in, , , are preset weights, and ; S305. Calculate the depth feature of the anchor sample before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows: ; S306. Calculate the distribution loss of the anchor sample after binarization: ; S307, perform weighted average calculation on the contrast loss calculated in step S301, the quantization loss calculated in step S305, and the distribution loss calculated in step S306 to obtain the anchor sample encoder model loss, which is formally expressed as follows: ; S308. Calculate the depth features of the negative sample set before binarization And the binary representation after binarization The mean square error is the quantization loss, which is formally expressed as follows: ; S309. Calculate the distribution loss of negative samples after binarization: ; S310, performing weighted average calculation on the contrast loss calculated in step S301, the quantization loss calculated in step S308, and the distribution loss calculated in step S309 to obtain the negative sample encoder model loss, which is formally expressed as follows: ; Step 4: Update the anchor sample encoder model, the positive sample encoder model, and the negative sample encoder model according to the gradient, calculate the standard encoder model parameters to be updated with the updated anchor sample encoder model parameters, the positive sample encoder model parameters, the negative sample encoder model parameters, and the current standard encoder model parameters and the preset weights, and finally update the standard encoder model; The step 4 comprises the following steps: Step 401: Based on the gradient, back-propagate the gradient in the positive sample encoder model, the negative sample encoder model and the anchor sample encoder model to update the positive sample encoder model, the negative sample encoder model and the anchor sample encoder model; Step 402: performing weighted average calculation on the updated positive sample encoder model parameters, negative sample encoder model parameters, anchor sample encoder model parameters and current standard encoder model parameters to obtain updated standard encoder model parameters and finally update the standard encoder model; Step 5: Repeat steps 2 to 4 until the predetermined training rounds are reached, then use the validation set to verify the retrieval performance of the standard encoder model, and save the best model among all current standard encoder models as the low-quality image representation model.
2. The low-quality image representation model training method with momentum weighted update according to claim 1, It is characterized in that The method further includes step 6: repeating steps 2 to 5 until a predetermined total number of training rounds is reached, thereby completing the training of the low-quality image representation model.
3. The low-quality image representation model training method with momentum weighted update according to claim 1, It is characterized in that In step 2, the training process of the anchor sample encoder model, the positive sample encoder model and the negative sample encoder model is independent training.
4. The low-quality image representation model training method with momentum weighted update according to any one of claims 1 to 3, It is characterized in that The standard encoder model, anchor sample encoder model, positive sample encoder model and negative sample encoder model are all composed of a deep encoder and a binary network, and have the same network structure and the same model parameters.
5. The low-quality image representation model training method with momentum weighted update according to claim 1, It is characterized in that After executing step 4, repeat steps 3 to 4, and clear the negative samples in the chain stack before each repetition until all negative samples in the self-supervised training sample unit are selected, and then continue to step 5.
6. The low-quality image representation model training method with momentum weighted update according to claim 5, It is characterized in that When performing weighted average calculation, the weight of the anchor sample encoder model parameters is greater than the weight of the positive sample encoder model parameters and the negative sample encoder model parameters.
7. A momentum-based weighted update low-quality image representation model training system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the computer program, the steps of the method described in any one of claims 1 to 6 are implemented.
