An unsupervised image hash retrieval method and system based on equal constraint clustering

Abstract

The present invention discloses an unsupervised image hash retrieval method and system based on equal-constrained clustering, comprising: inputting a database image into a hash feature extraction network model, obtaining a binary hash code of the database image and constructing a hash code database; inputting a to-be-retrieved image into a trained hash feature extraction network model, obtaining a hash code of the to-be-retrieved image, calculating the Hamming distance between the hash code of the to-be-retrieved image and all the binary hash codes in the hash code database, and taking the database image with the smallest Hamming distance as the retrieval result of the to-be-retrieved image. The present invention uses equal-constrained clustering to cluster unlabeled training images, effectively avoiding the empty clustering and cluster imbalance problems faced by the K-Means clustering algorithm, and improving the accuracy of hard-pseudo-label annotation; using denoised soft-pseudo-labels to supervise hash feature extraction network training, making full use of soft-pseudo-labels to mine image class relationships, avoiding overfitting caused by hard-pseudo-labels, and improving hash coding accuracy.

Description

Technical Field

[0001] The invention relates to an unsupervised image hash retrieval method based on equal amount constrained clustering and an unsupervised image hash retrieval system based on equal amount constrained clustering, belonging to the technical field of computer vision and image retrieval. Background Art

[0002] With the rapid development of deep learning technology, deep neural networks have been widely used in various fields of computer vision. In image retrieval tasks, image hash features extracted by deep neural networks have the advantages of high storage efficiency and fast query speed, which has attracted extensive attention from researchers. At present, image hash retrieval methods are mainly based on supervised deep learning methods for representation learning. This type of method combines the advantages of deep learning and hash learning, improving retrieval accuracy while maintaining retrieval efficiency. However, supervised deep hashing methods rely on a large amount of accurately labeled data, and obtaining this data will consume a lot of human resources.

[0003] In the real world, the amount of unlabeled data far exceeds that of labeled data. How to use massive unlabeled data to obtain "knowledge" has become a very challenging topic. Unlike supervised deep hash learning methods, unsupervised deep hash learning methods do not require manually labeled data. They can directly learn compact hash features from unlabeled data for hash retrieval, greatly reducing the cost of hash learning. Therefore, unsupervised deep hash learning has become a new research hotspot.

[0004] Currently, most of the unsupervised deep hash learning methods for images learn hash codes through proxy tasks. For example, a Chinese patent with application number 201810134469.2 and publication number CN108491430A proposes an unsupervised hash retrieval method based on clustering feature directions. This method clusters features through the K-Means clustering algorithm to achieve feature dimensionality reduction, and further clusters each dimension of the new features and implements binary encoding. However, due to the lack of effective constraints in the K-Means algorithm, it will face the problems of empty clusters and unbalanced clusters, and this type of method does not perform robust learning processing, which will result in the resulting hash code being not robust enough, thereby affecting the accuracy of hash retrieval.

[0005] In summary, although the existing unsupervised deep hash retrieval methods can achieve efficient hash retrieval, they are limited by the performance of their own annotation methods, so there is still a certain gap between the performance of deep unsupervised hash methods and deep supervised hash methods. How to achieve high-precision data annotation and improve the robustness of deep hash networks to label noise has become the research focus of unsupervised hash learning methods. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide an unsupervised image hash retrieval method and system based on equal-constrained clustering to solve the problems of empty clusters, unbalanced clusters and the problem that the hash learning method is not robust enough faced by the existing clustering-based unsupervised deep hash image retrieval methods.

[0007] To achieve the above object, the present invention provides an unsupervised image hash retrieval method based on equal constraint clustering, comprising:

[0008] Get the image to be retrieved;

[0009] Input the database image into the trained hash feature extraction network model to obtain the binary hash code of the database image. The binary hash codes of all database images constitute a hash code database.

[0010] The image to be retrieved is input into the trained hash feature extraction network model to obtain the hash code of the image to be retrieved, the Hamming distance between the hash code of the image to be retrieved and all the binary hash codes in the hash code database is calculated, and the database image with the smallest Hamming distance is taken as the retrieval result of the image to be retrieved.

[0011] Prioritize, train and obtain a hash feature extraction network model, including:

[0012] Get unlabeled training images;

[0013] The unlabeled training images are labeled using a soft pseudo label labeling method to obtain soft pseudo labels of the unlabeled training images, and a soft pseudo label denoising method is used to remove some noise labels of the soft pseudo labels;

[0014] Based on the obtained soft pseudo labels with noise removed, the preset loss function Λ S The parameters of the pre-built hash feature extraction network are iteratively updated using gradient update. After the preset number of iterative updates is met, the final hash feature extraction network model is obtained.

[0015] Preferably, the soft fake label marking method includes:

[0016] Use pre-trained deep networks to extract deep features from unlabeled training images;

[0017] The deep features are clustered using the equal-constrained clustering algorithm, and hard pseudo labels are annotated on the unlabeled training images according to the clusters where the deep features are located;

[0018] Based on the obtained hard false labels, the unlabeled training images are annotated with soft false labels using the trained soft false label annotation network model to obtain the soft false labels of the unlabeled training images.

[0019] Prioritize training to obtain soft pseudo-label annotation network models, including:

[0020] obtaining an unlabeled training dataset including unlabeled training images with hard pseudo labels;

[0021] The unlabeled training dataset is input into the pre-built soft pseudo-label annotation network, and the preset loss function Λ H Gradient update is used to iteratively update the parameters of the pre-built soft pseudo label annotation network. After the preset number of iterative updates is met, the final soft pseudo label annotation network model is output.

[0022] Preferably, the equal amount constraint clustering algorithm is a K-Means clustering algorithm with cluster equal amount constraint added, and the objective function J of the equal amount constraint clustering algorithm is:

[0023]

[0024] Where f represents the deep features of the unlabeled training image extracted using the pre-trained deep network, C i represents the i-th cluster, u i represents the cluster center of the i-th cluster, N represents the number of unlabeled training images; k represents the number of unlabeled training image categories, i∈[1,k].

[0025] Preferably, the preset loss function Λ H for:

[0026]

[0027] In the formula, represents the hard pseudo labels of the unlabeled training images; Indicates a soft false label.

[0028] Preferably, the preset loss function Λ S for:

[0029]

[0030] In the formula, represents the predicted probability distribution of the hash feature extraction network for the unlabeled training images, k represents the number of unlabeled training image categories, i∈[1,k].

[0031] Preferably, the gradient is updated as:

[0032] θ=θ 0 -η·g'(θ 0 ), where θ 0 represents the network parameters before gradient update, θ represents the network parameters after gradient update; η represents the learning rate, which ranges from [0,1], g'(θ0 ) represents the hash feature extraction network or the soft pseudo label annotation network in θ 0 The gradient at .

[0033] Preferably, an acquisition module is used to acquire the image to be retrieved;

[0034] A hash code database acquisition module is used to input the database image into the trained hash feature extraction network model to obtain the binary hash code of the database image. The binary hash codes of all database images constitute the hash code database.

[0035] A retrieval module is used to input the image to be retrieved into the trained hash feature extraction network model to obtain the hash code of the image to be retrieved;

[0036] A Hamming distance calculation module, used for calculating the Hamming distance between the hash code of the image to be retrieved and all the binary hash codes in the hash code database;

[0037] The result determination module is used to take the database image with the smallest Hamming distance as the retrieval result of the image to be retrieved.

[0038] An unsupervised image hash retrieval device based on equal-constrained clustering, comprising a processor and a storage medium;

[0039] The storage medium is used to store instructions;

[0040] The processor is used to operate according to the instructions to execute the steps of the above method.

[0041] A computer-readable storage medium stores a computer program, which implements the steps of any of the above methods when executed by a processor.

[0042] The beneficial effects achieved by the present invention are:

[0043] The present invention inputs the image to be retrieved into a pre-trained hash feature extraction network to obtain a hash code of the image to be retrieved; calculates the Hamming distance between the hash code of the image to be retrieved and the hash code in the hash code database, and sorts them from small to large according to the Hamming distance, and outputs the top-ranked database image as the retrieval result;

[0044] The present invention uses equal-constrained clustering to cluster unlabeled training images, which can effectively avoid the empty cluster and cluster imbalance problems faced by the K-Means clustering algorithm, thereby improving the accuracy of hard-pseudo label annotation; using hard-pseudo labels to train soft-pseudo label annotation networks, the soft-pseudo label annotation network can be adapted to the data and the soft-pseudo label annotation effect can be improved; using a soft-pseudo label denoising method to remove noise labels in soft-pseudo labels, the accuracy of soft-pseudo labels can be further improved; using denoised soft-pseudo labels to supervise hash feature extraction network training, the soft-pseudo labels can be fully utilized to mine the relationship between image classes, avoid overfitting caused by "one-hot" hard-pseudo labels, improve hash coding accuracy, and better complete image retrieval tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the method of the present invention;

[0046] Figure 2 It is a flow chart of constructing a hash code database in the present invention;

[0047] Figure 3 It is a schematic diagram of the search of the method of the present invention. DETAILED DESCRIPTION

[0048] The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and are not intended to limit the protection scope of the present invention.

[0049] Embodiment 1

[0050] like Figure 3 As shown, an unsupervised image hash retrieval method based on equal constraint clustering includes:

[0051] Input the database image into the trained hash feature extraction network model to obtain the binary hash code of the database image. The binary hash codes of all database images constitute a hash code database.

[0052] The image to be retrieved is input into the trained hash feature extraction network model to obtain the hash code of the image to be retrieved, the Hamming distance between the hash code of the image to be retrieved and all the binary hash codes in the hash code database is calculated, and the database image with the smallest Hamming distance is taken as the retrieval result of the image to be retrieved.

[0053] like Figure 1 As shown, the hash feature extraction network model is trained, including:

[0054] Get an unlabeled training image set X;

[0055] The unlabeled training image set X includes unlabeled training images for training a hash feature extraction network;

[0056] The unlabeled training image set X is labeled using the soft pseudo label labeling method to obtain the soft pseudo labels of the unlabeled training images, and the soft pseudo label denoising method is used to remove some noise labels;

[0057] Based on the obtained unlabeled training image denoising soft pseudo-label, the preset loss function Λ S The pre-built hash feature extraction network H(.) is iteratively updated with gradient update to update its parameters. After the preset number of iterative updates is met, the final hash feature extraction network model is obtained.

[0058] Input the database image into the hash feature extraction network model H(.) that has completed parameter iterative update to obtain a hash code database;

[0059] The image q to be retrieved is input into the hash feature extraction network model H(.) to obtain the hash code H(q) of the image to be retrieved, the Hamming distance between H(q) and the binary hash codes in all hash code databases is calculated, and the database images with the highest ranking are sorted from small to large according to the Hamming distance, and the top-ranked database images are output as the retrieval results.

[0060] The soft fake label marking method comprises:

[0061] Use pre-trained deep networks to extract deep features from unlabeled training images;

[0062] The acquired deep features are clustered using the equal constraint clustering algorithm, and hard false labels are annotated on the image according to the clusters where the deep features are located;

[0063] Based on the obtained hard false labels of the images, the unlabeled training images are annotated with soft false labels using the trained soft false label annotation network model to obtain the soft false labels of the unlabeled training images.

[0064] Train the soft pseudo-label annotation network model, including:

[0065] obtaining an unlabeled training dataset including unlabeled training images with hard pseudo labels;

[0066] The unlabeled training dataset is input into the pre-built soft pseudo-label annotation network, and the preset loss function Λ H Gradient update is used to iteratively update the parameters of the pre-built soft pseudo label annotation network. After the preset number of iterative updates is met, the final soft pseudo label annotation network model is output.

[0067] The network structure of the soft pseudo-label annotation network model is an existing technology, and an existing image deep feature extraction model such as ResNet can be used. The internal network structure of the hash feature extraction network model is an existing technology.

[0068] The specific steps of obtaining the hash feature extraction network H(.) and the hash code database include:

[0069] Step 1: Obtain unlabeled training image data and use the pre-trained deep network to extract deep features of the image.

[0070] Unlabeled training images and database images can be obtained through web crawlers, video extraction, etc., which are low-cost and have basically no restrictions on acquisition conditions. The pre-trained deep network can select a ResNet deep network pre-trained on the ImageNet dataset, and use the pre-trained ResNet deep network to extract 2048-dimensional deep features. For related methods of ResNet deep network structure, see He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[J]. IEEE, 2016.

[0071] Step 2: If Figure 2 As shown, the equal-constrained clustering algorithm is used to cluster the deep features of the unlabeled training image, and the cluster number of the deep feature is output as the hard pseudo label of the image.

[0072] The equal-constrained clustering algorithm adds clustering equal-constraints based on the K-Means clustering algorithm, which requires that the sample size of each cluster is equal during the clustering process. The objective function J of the equal-constrained clustering algorithm is:

[0073]

[0074] (1) where f represents the image depth feature extracted using the pre-trained deep network, C i represents the i-th cluster, u i represents the cluster center of the i-th cluster, N represents the number of images, and k represents the number of image categories.

[0075] After clustering is completed, the cluster number of the deep feature can be output as the hard pseudo label of the unlabeled training image. The hard pseudo label is a "one-hot" vector with a length equal to the number of image categories, where the position element corresponding to the cluster number of the unlabeled training image deep feature is "1", and the other position elements are "0".

[0076] Step 3: If Figure 2 As shown in Figure 1, using hard false labels to train the soft false label annotation network can make the soft false label annotation network adapt to the data and obtain higher quality soft false labels. H for:

[0077]

[0078] In the formula, represents the hard pseudo labels of the unlabeled training images; It represents the predicted probability distribution of the soft pseudo label annotation network for the unlabeled training images. The number of training times should not be too many, preferably 2 epochs, to better preserve the inter-class relationship in the predicted probability distribution. After the soft pseudo label annotation network training is completed, the unlabeled training images can be input into the soft pseudo label annotation network for prediction, and the predicted probability distribution of the unlabeled training images can be obtained as soft pseudo labels.

[0079] Step 4: Figure 2 As shown, the soft false label denoising method can be used to remove the label noise in the soft false label and improve the accuracy of the soft false label. The soft false label denoising method includes distance denoising and confidence denoising. The distance denoising will filter out the soft false labels of the unlabeled training images far away from the cluster center feature, and retain the soft false labels of the unlabeled training images closest to the cluster center feature of each cluster according to the ratio r; the confidence denoising is performed after the distance denoising is performed, and the confidence denoising will filter out the soft false labels with a confidence less than the threshold e.

[0080] Step 5: Figure 2 As shown in Figure 1, using denoised soft pseudo labels to train the hash feature extraction network can fully explore the inter-class relationship of the image and avoid overfitting caused by "one-hot" hard pseudo labels. The loss function Λ used in training S for:

[0081]

[0082] (3) In the formula, Represents the predicted probability distribution of the hash feature extraction network for unlabeled training images.

[0083] When training the network, the gradient update satisfies:

[0084] θ=θ 0 -η·g'(θ 0 ) (4)

[0085] (4) In the formula, θ 0 and θ represent the network parameters before and after gradient update, respectively; η represents the learning rate, which ranges from [0,1], g'(θ 0 ) represents the hash feature extraction network or the soft pseudo label annotation network in θ 0 The gradient at .

[0086] Step 6: Input the database image data into the trained hash feature extraction network H(.) to obtain the binary hash code of the database image. The binary hash codes of the database image are combined to generate a hash code database. The binary hash code b is:

[0087] b=tanh(Φ(x)) (5)

[0088] (5)In the formula, x represents the database image and Φ represents the hash feature extraction network.

[0089] Database images, unlabeled training images, and unlabeled image data can be collected online through web crawlers.

[0090] The network structure of the soft pseudo-label annotation network model can adopt the existing image deep feature extraction model, such as ResNet.

[0091] The hard false label is a "one-hot" vector with a length equal to the number of unlabeled training image categories, and the corresponding position element of the number of the cluster where the deep feature is located is "1", and the remaining position elements are "0".

[0092] Embodiment 2

[0093] This embodiment provides an unsupervised image hash retrieval system based on equal constraint clustering, including:

[0094] Acquisition module: used to obtain the image q to be retrieved;

[0095] Retrieval module: used to input the image q to be retrieved into the pre-obtained hash feature extraction network H(.) to obtain the hash code H(q) of the image to be retrieved; calculate the Hamming distance between H(q) and the binary hash code of the database image in the hash code database and sort them. The smaller the Hamming distance, the higher the retrieval ranking. The top-ranked database image is output as the retrieval result.

[0096] Among them, the retrieval module includes a construction module: used to construct a hash feature extraction network H(.) and a hash code database.

[0097] The building blocks include:

[0098] Deep feature extraction component: used to extract deep features of unlabeled training images using pre-trained deep networks;

[0099] Hard false label annotation component: used to cluster the acquired unlabeled training image deep features using the equal constraint clustering algorithm, and annotate the unlabeled training image with hard false labels according to the clusters where the unlabeled training image deep features are located;

[0100] Soft pseudo-label annotation component: Based on the obtained hard pseudo-label of unlabeled training images, the preset loss function Λ H The pre-built soft pseudo label annotation network is updated iteratively using gradient update. The unlabeled training images are annotated with soft pseudo labels using the updated soft pseudo label annotation network to obtain the soft pseudo labels of the unlabeled training images.

[0101] Soft false label denoising component: denoise the obtained soft false labels of unlabeled training images using the soft false label denoising method, remove the soft false labels corresponding to features with confidence values ​​lower than the threshold and far from the cluster center, and obtain denoised soft false labels;

[0102] Training component: used to denoise the soft pseudo labels obtained, using the preset loss function Λ S Gradient update is used to update the parameters of the pre-built hash feature extraction network H(.);

[0103] Output component: used to input the database image into the hash feature extraction network H(.) that completes parameter update, and obtain the binary hash code of the database image. The hash codes of all database images constitute the hash code database.

[0104] Embodiment 3

[0105] The embodiment of the present invention provides an unsupervised image hash retrieval device based on equal-constrained clustering, including a processor and a storage medium;

[0106] The storage medium is used to store instructions;

[0107] The processor is used to operate according to the instructions to execute the steps of the method described in embodiment 1.

[0108] Embodiment 4

[0109] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the first embodiment are implemented.

[0110] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0111] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0112] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0113] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An unsupervised image hash retrieval method based on equal constraint clustering, characterized in that: include: Get the image to be retrieved; Input the database image into the trained hash feature extraction network model to obtain the binary hash code of the database image. The binary hash codes of all database images constitute a hash code database. Input the image to be retrieved into the trained hash feature extraction network model to obtain the hash code of the image to be retrieved, calculate the Hamming distance between the hash code of the image to be retrieved and all the binary hash codes in the hash code database, and take the database image with the smallest Hamming distance as the retrieval result of the image to be retrieved; Among them, the hash feature extraction network model is trained, including: Get unlabeled training images; The unlabeled training images are labeled using a soft pseudo label labeling method to obtain soft pseudo labels of the unlabeled training images, and a soft pseudo label denoising method is used to remove some noise labels of the soft pseudo labels; Based on the obtained soft pseudo labels with noise removed, the preset loss function The parameters of the pre-built hash feature extraction network are iteratively updated using gradient update. After the preset number of iterative updates is met, the final hash feature extraction network model is obtained. The soft false label marking method comprises: Use pre-trained deep networks to extract deep features from unlabeled training images; The deep features are clustered using the equal-constrained clustering algorithm, and hard pseudo labels are annotated on the unlabeled training images according to the clusters where the deep features are located; Based on the obtained hard false labels, the unlabeled training images are labeled with soft false labels using the trained soft false label labeling network model to obtain soft false labels for the unlabeled training images; The equal amount constraint clustering algorithm adds cluster equal amount constraint to the K-Means clustering algorithm. The objective function of the equal amount constraint clustering algorithm is for: ， In the formula, represents the deep features of unlabeled training images extracted using a pre-trained deep network, Indicates Clusters, Indicates The cluster centers of the clusters, represents the number of unlabeled training images; Represents the number of unlabeled training image categories, i∈[1,k].

2. The unsupervised image hash retrieval method based on equal-constrained clustering according to claim 1 is characterized in that: Train the soft pseudo-label annotation network model, including: obtaining an unlabeled training dataset including unlabeled training images with hard pseudo labels; The unlabeled training dataset is input into the pre-built soft pseudo-label annotation network, and the preset loss function Gradient update is used to iteratively update the parameters of the pre-built soft pseudo label annotation network. After the preset number of iterative updates is met, the final soft pseudo label annotation network model is output.

3. The unsupervised image hash retrieval method based on equal-constrained clustering according to claim 2 is characterized in that: The preset loss function for: ， In the formula, represents the hard pseudo labels of the unlabeled training images; Indicates a soft fake label; represents the number of unlabeled training image categories, i∈[1,k].

4. The unsupervised image hash retrieval method based on equal-constrained clustering according to claim 3 is characterized in that: The preset loss function for: ， In the formula, represents the predicted probability distribution of the hash feature extraction network for unlabeled training images, Indicates a soft fake label; Represents the number of unlabeled training image categories, i∈[1,k].

5. The unsupervised image hash retrieval method based on equal-constrained clustering according to claim 1, characterized in that: The gradient update is: ， In the formula, represents the network parameters before gradient update, Represents the network parameters after gradient update; Represents the learning rate, the value range is [0,1], Indicates that the hash feature extraction network or the soft pseudo label annotation network is The gradient at .

6. An unsupervised image hash retrieval system based on equal constraint clustering, characterized in that: An acquisition module, used for acquiring the image to be retrieved; A hash code database acquisition module is used to input the database image into the trained hash feature extraction network model to obtain the binary hash code of the database image. The binary hash codes of all database images constitute the hash code database. A retrieval module is used to input the image to be retrieved into the trained hash feature extraction network model to obtain the hash code of the image to be retrieved; A Hamming distance calculation module, used for calculating the Hamming distance between the hash code of the image to be retrieved and all the binary hash codes in the hash code database; A result determination module, used for taking the database image with the smallest Hamming distance as the retrieval result of the image to be retrieved; Among them, the hash feature extraction network model is trained, including: Get unlabeled training images; The unlabeled training images are labeled using a soft pseudo label labeling method to obtain soft pseudo labels of the unlabeled training images, and a soft pseudo label denoising method is used to remove some noise labels of the soft pseudo labels; Based on the obtained soft pseudo labels with noise removed, the preset loss function The parameters of the pre-built hash feature extraction network are iteratively updated using gradient update. After the preset number of iterative updates is met, the final hash feature extraction network model is obtained. The soft false label marking method comprises: Use pre-trained deep networks to extract deep features from unlabeled training images; The deep features are clustered using the equal-constrained clustering algorithm, and hard pseudo labels are annotated on the unlabeled training images according to the clusters where the deep features are located; Based on the obtained hard false labels, the unlabeled training images are labeled with soft false labels using the trained soft false label labeling network model to obtain soft false labels for the unlabeled training images; The equal amount constraint clustering algorithm adds cluster equal amount constraint to the K-Means clustering algorithm. The objective function of the equal amount constraint clustering algorithm is for: ， In the formula, represents the deep features of unlabeled training images extracted using a pre-trained deep network, Indicates Clusters, Indicates The cluster centers of the clusters, represents the number of unlabeled training images; represents the number of unlabeled training image categories, i∈[1,k].

7. An unsupervised image hash retrieval device based on equal constraint clustering, characterized in that: including processor and storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method described in any one of claims 1 to 5 are implemented.

Images