Multi-granularity perception and prototype-driven adaptive recognition method for ultra-high-definition image domain

Abstract

This invention belongs to the field of domain adaptation technology for ultra-high-definition (UHD) images. It proposes a multi-granularity sensing and prototype-driven UHD image domain adaptive recognition method, comprising the following steps: First, acquiring a cross-image domain conventional image dataset and an UHD image dataset, and preprocessing both datasets to obtain a preprocessed batch training image dataset; second, constructing and initializing a multi-granularity feature extractor and a feature classifier; then, training an unsupervised domain adaptive recognition model for UHD images based on the batch training image dataset and combining the multi-granularity feature extractor and the feature classifier; finally, inputting the UHD image to be recognized into the unsupervised domain adaptive recognition model for classification to obtain the recognition result. This invention solves the problems of neglecting fine-grained features in UHD images and the difficulties, limited quantity, and high computational resource consumption of UHD image domain annotation.

Description

Technical Field

[0001] This invention relates to the field of domain adaptation technology for ultra-high-definition images, and particularly to a domain adaptive recognition method for ultra-high-definition images based on multi-granularity perception and prototype-driven approaches. Background Technology

[0002] With the rapid development of machine vision technology, ultra-high-definition (UHD) imaging technology has begun to be applied in fields such as surveillance and reconnaissance. However, due to the difficulty in acquiring UHD images and the scarcity of well-labeled data, it is difficult to directly train a neural network model with strong generalization ability using a small number of well-labeled UHD images. Furthermore, if unsupervised learning or a large number of UHD images with potentially noisy labels are used for training, the prediction accuracy and robustness of the resulting deep neural network model may not meet real-world requirements. Therefore, how to utilize unsupervised domain adaptation techniques to transfer supervised knowledge from currently mature conventional image recognition to the UHD image domain, thereby improving the generalization ability and robustness of UHD image recognition models, is a key solution for improving the performance of UHD image recognition.

[0003] Existing ultra-high-definition image recognition technologies directly employ deep models such as convolutional neural networks and Transformers to extract features from a small number of fully annotated ultra-high-definition images, and then input these image features into a classifier to obtain the final recognition result. Alternatively, ultra-high-definition images are downsampled to become regular images, and then fed into a deep neural network for regular images for learning and recognition.

[0004] Furthermore, some unsupervised domain adaptation methods have begun to be applied to this field. In these methods, the labeled source domain corresponds to the domain of regular images, while the unlabeled domain corresponds to the domain of ultra-high-definition images. Traditional unsupervised domain adaptation methods typically treat regular and ultra-high-definition images as belonging to different data distributions, and then eliminate the differences between sample pairs at the instance level—a process known as domain shift. With the development of deep learning, more methods employ adversarial techniques to address the impact of domain shift. They use ultra-high-definition images as an adversarial set against regular images, striving to reduce the differences between the two at the pixel level. In addition, self-supervised domain adaptation methods directly classify ultra-high-definition images using pre-trained models of regular images to obtain pseudo-labels, which are then used as supervisory signals for learning, achieving recognition of ultra-high-definition images.

[0005] Therefore, existing ultra-high-definition image recognition methods are still limited to mature conventional image recognition techniques or directly applying unsupervised adaptive method templates, ignoring the unique characteristics of ultra-high-definition images themselves. Generally speaking, these existing ultra-high-definition image recognition technologies share some common limitations:

[0006] (1) The vast amount of fine-grained information contained in ultra-high-definition images has not been well utilized, and has even been directly discarded by some existing methods.

[0007] (2) It ignores the characteristics of ultra-high-definition images being difficult to annotate and having a small number of images, making it difficult to achieve strong generalization recognition performance through direct training.

[0008] (3) Ultra-high-definition images are large in size, and both instance-based and adversarial deep learning methods face huge time and space computational costs. Summary of the Invention

[0009] The purpose of this invention is to provide an ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven approach. This method can fully grasp the characteristics of ultra-high-definition images, such as rich details, sparse annotations, and large image size, achieving full perception of fine-grained features and solving the problem of neglecting these features. Simultaneously, it employs a large number of labeled conventional images and a small number of unlabeled ultra-high-definition images to achieve unsupervised domain adaptive recognition. Lightweight prototype learning techniques are used to transfer supervised knowledge from conventional images to the ultra-high-definition image domain, achieving more accurate ultra-high-definition image recognition and solving the problems of difficult and limited annotation resources and high computational resource consumption in the ultra-high-definition image domain.

[0010] The technical solution adopted by this invention to solve its technical problem is as follows:

[0011] A multi-granularity sensing and prototype-driven ultra-high-definition image domain adaptive recognition method includes the following steps:

[0012] Obtain regular image datasets and ultra-high-definition image datasets across image domains, and preprocess the regular image datasets and ultra-high-definition image datasets to obtain preprocessed batch training image datasets;

[0013] Construct and initialize a multi-granularity feature extractor and a feature classifier;

[0014] Based on a batch training image dataset, an unsupervised adaptive recognition model for ultra-high-definition images is obtained by combining a multi-granularity feature extractor and a feature classifier.

[0015] The ultra-high-definition image to be identified is input into an unsupervised adaptive recognition model based on ultra-high-definition images for classification, and the recognition result of the ultra-high-definition image is obtained.

[0016] As a further optimization, the step of obtaining a regular image dataset and an ultra-high-resolution image dataset across image domains, and preprocessing the regular image dataset and the ultra-high-resolution image dataset to obtain a preprocessed batch training image dataset includes the following steps:

[0017] Obtain a regular image dataset with accurate labels, which contains regular natural images of various scenes in production and daily life;

[0018] Obtain an unlabeled ultra-high-resolution image dataset;

[0019] Data cleaning was performed on the two image datasets mentioned above to remove erroneous and noisy images, resulting in a regular image dataset. and ultra-high-definition image datasets

[0020] regular image datasets and ultra-high-definition image datasets The training data is divided and organized to obtain a batch training image dataset.

[0021] As a further optimization, the construction and initialization of the multi-granularity feature extractor and feature classifier includes the following steps:

[0022] Establish and initialize the Faster RCNN network framework with pre-trained parameters to obtain the object detection model.

[0023] Establish and initialize the ResNet-18 network framework with pre-trained parameters to obtain the backbone network model.

[0024] A single-layer convolutional neural network framework is established and randomly initialized to obtain a fine-grained attention network. A single-layer fully connected neural network framework was established to obtain a feature-granularity attention network.

[0025] Combined target detection model Backbone network model Fine-grained attention network and feature-granularity attention network Obtain a multi-granularity feature extractor

[0026] A multi-layer fully connected neural network framework is established and randomly initialized to obtain the feature classifier f.

[0027] As a further optimization, the ultra-high-definition image unsupervised adaptive recognition model, trained based on batch training image datasets and combined with multi-granularity feature extractors and feature classifiers, includes the following steps:

[0028] Batch training image dataset Input to multi-granularity feature extractor In the process, multi-granularity features are extracted from each batch of images.

[0029] Using a regular image as the source domain, and incorporating multi-granular features from the source domain... Based on the image category information and the K-Means clustering algorithm, the prototype center c of the source domain is calculated. s ;

[0030] Using ultra-high-resolution images as the target domain, and incorporating multi-granularity features of the target domain... The K-Means clustering algorithm is used to obtain the prototype centers c of the target domain. t ;

[0031] Utilizing the multi-granularity features of source domain images and the prototype center c of the source domain s And utilizing the multi-granularity features of the target domain image and the prototype center c of the target domain t In each domain, the domain cohesion loss is calculated.

[0032] Multi-granularity features of joint source domain images and the prototype center c of the source domain s and multi-granular features of the target domain image and the prototype center c of the target domain t The inter-domain coupling loss is calculated between domains.

[0033] Multi-granular features of the source domain image Multi-granularity features of the target domain image The input is fed into the feature classifier to obtain the feature classification offset loss.

[0034] According to the cohesion loss in the domain Inter-domain coupling loss and feature classification offset loss The total loss of the entire model optimization is calculated.

[0035] Calculate total loss The gradient is used to backpropagate and optimize the parameters of the entire model. When the total loss meets the convergence condition, the trained ultra-high-definition image unsupervised adaptive recognition model is obtained.

[0036] As a further optimization, the calculation formula for the multi-granularity features is as follows:

[0037]

[0038]

[0039]

[0040] in, It is a multi-granularity feature set of batch images. It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. It is a multi-granularity feature extractor. and These are batch training datasets for source domain images and target domain images, respectively.

[0041] As a further optimization, the prototype center c of the source domain s The calculation formula is:

[0042]

[0043] The prototype center c of the target domain t The calculation formula is:

[0044]

[0045] Among them, c s This represents the prototype center of the constructed source domain. This represents the cluster centers obtained after performing K-Means clustering on the features of the source domain. Let represent the category center of each category in the source domain, e represent a weight hyperparameter used to balance the proportion of cluster centers and category centers in the construction of prototype centers, and c represent the category center. t Indicates the prototype center of the target domain being constructed, using To construct the prototype center, This represents the cluster centers obtained after K-Means clustering of the target domain features.

[0046] As a further optimization, the cohesion loss in the domain The expression is:

[0047]

[0048] in, The cohesion loss function in the domain, n s The total amount of source domain image data, n t The total amount of image data in the target domain, i and j are indicator variables for iteratively calculating the loss function of the image data, and CE refers to the cross-entropy loss function. This represents the i-th feature extracted from the source domain image. This represents the extracted j-th target domain image feature. It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. Representative target domain features The index of the nearest cluster center;

[0049] The inter-domain coupling loss The expression is:

[0050]

[0051] in, This represents the inter-domain coupling loss function, and entropy refers to the calculation of information entropy. The cross-domain prototype center representing the feature cluster, Representative source domain features and target domain features The latest cross-domain prototype center c g The index;

[0052] The feature classification offset loss The expression is:

[0053]

[0054] in, This represents the feature classification offset loss. represents the category label corresponding to the source domain feature, and f represents the feature classifier that has been constructed. It is a function representation for classifying features.

[0055] As a further optimization, the expression for the total loss is:

[0056]

[0057] in, Represents the total loss function. α, β, and γ represent the domain cohesion loss function, the inter-domain coupling loss function, and the feature classification offset loss function, respectively, and represent the weight values ​​of the corresponding function terms.

[0058] As a further optimization, the total loss satisfying the convergence condition means that the total loss of 5 consecutive batches satisfies the following:

[0059]

[0060] in, For the total loss of the ep-th batch, Let be the total loss for the (ep-1)th batch, and |·| be the absolute value function.

[0061] The beneficial effects of this invention are as follows: Through the aforementioned multi-granularity perception and prototype-driven ultra-high-definition image domain adaptive recognition method, based on a multi-granularity attention perception mechanism, image convolution operations, pre-trained image encoders, and attention mechanisms are utilized to achieve ultra-high-definition image feature extraction that integrates local and global information, endowing the model with the ability to comprehensively extract fine-grained and overall features of ultra-high-definition images. Furthermore, this invention employs an unsupervised domain adaptive ultra-high-definition image recognition solution based on conventional images as the source domain, enabling the transfer of extensive supervised knowledge from conventional images to the ultra-high-definition image domain, thereby improving the model's performance and generalization ability in ultra-high-definition image recognition. Additionally, this invention uses prototype learning as the driving technology for domain adaptation, achieving cohesive learning within the domain and coupling learning between domains through a "one-to-many" paradigm of prototype center-common representation, thereby improving the model's discriminative power and eliminating domain bias with lower computational cost. Attached Figure Description

[0062] Figure 1 This is a flowchart of the ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0064] Example

[0065] This embodiment provides a multi-granularity sensing and prototype-driven ultra-high-definition image domain adaptive recognition method, the flowchart of which can be found in [link to flowchart]. Figure 1 The method includes the following steps:

[0066] S1. Obtain regular image datasets and ultra-high-definition image datasets across image domains, and preprocess the regular image datasets and ultra-high-definition image datasets to obtain preprocessed batch training image datasets.

[0067] S2. Construct and initialize the multi-granularity feature extractor and feature classifier;

[0068] S3. Based on the batch training image dataset, an unsupervised adaptive recognition model for ultra-high-definition images is trained by combining a multi-granularity feature extractor and a feature classifier.

[0069] S4. Input the ultra-high-definition image to be identified into the ultra-high-definition image unsupervised domain adaptive recognition model for classification, and obtain the recognition result of the ultra-high-definition image.

[0070] In this embodiment, firstly, based on a multi-granularity attention perception mechanism, image convolution operations, pre-trained image encoders, and attention mechanisms are used to achieve ultra-high-definition image feature extraction that integrates local and global information, giving the model the ability to comprehensively extract fine-grained and overall features of ultra-high-definition images. Secondly, by adopting an unsupervised domain-adaptive ultra-high-definition image recognition solution based on conventional images as the source domain, the extensive supervised knowledge in conventional images can be transferred to the ultra-high-definition image domain, improving the model's performance and generalization ability in ultra-high-definition image recognition. Then, prototype learning is used as the driving technology for domain adaptation, and a "one-to-many" paradigm of prototype center-common representation is used to achieve cohesive learning in the domain and coupling learning between domains, thereby improving the model's discriminative power and eliminating domain bias with lower computational cost.

[0071] It should be noted that, in order to provide training data support and foundation for the subsequent implementation of unsupervised domain adaptive recognition technology for ultra-high-definition images, step S1 of this embodiment, which involves obtaining a conventional image dataset and an ultra-high-definition image dataset across image domains, and preprocessing the conventional image dataset and the ultra-high-definition image dataset to obtain a preprocessed batch training image dataset, may include the following steps:

[0072] S101. Obtain a regular image dataset with accurate labels, which contains regular natural images of various scenes in production and life.

[0073] S102. Obtain an unlabeled ultra-high-resolution image dataset;

[0074] S103. Perform data cleaning on the two image datasets above, filtering out erroneous and noisy image data to obtain a regular image dataset. and ultra-high-definition image datasets

[0075] S104, Transfer the regular image dataset and ultra-high-definition image datasets The training data is divided and organized to obtain a batch training image dataset.

[0076] It should be noted that by building specific multi-granularity feature extractors and feature classifiers, a model framework is established for subsequent multi-granularity feature perception and accurate recognition and classification of regular images and ultra-high-definition images. In step S2 of this embodiment, the construction and initialization of the multi-granularity feature extractor and feature classifier may include the following steps:

[0077] S201. Establish and initialize the Faster RCNN network framework with pre-trained parameters to obtain the object detection model.

[0078] S202. Establish and initialize the ResNet-18 network framework with pre-trained parameters to obtain the backbone network model.

[0079] S203. Establish and randomly initialize a single-layer convolutional neural network framework to obtain a fine-grained attention network. A single-layer fully connected neural network framework was established to obtain a feature-granularity attention network.

[0080] S204, Combined Target Detection Model Backbone network model Fine-grained attention network and feature-granularity attention network Obtain a multi-granularity feature extractor

[0081] S205. Establish and randomly initialize a multi-layer fully connected neural network framework to obtain the feature classifier f.

[0082] Furthermore, in order to extract fine-grained dimensional features and feature dimensional characteristics from the acquired trained image data through the constructed multi-granularity feature extractor, and fuse them into an image depth representation to fully explore every detail and feature of the image, this embodiment can promote the alignment of the conventional image domain to the ultra-high-definition image domain and the accurate recognition of ultra-high-definition images with less computation by using prototype-based supervision of domain cohesion loss and inter-domain coupling loss, thereby improving the unsupervised domain adaptive recognition capability of ultra-high-definition images. Therefore, in step S3 of this embodiment, the step of training an ultra-high-definition image unsupervised domain adaptive recognition model based on a batch training image dataset, combined with a multi-granularity feature extractor and a feature classifier, may include the following steps:

[0083] S301, batch training image dataset Input to multi-granularity feature extractor In the process, multi-granularity features are extracted from each batch of images.

[0084] S302. Use a regular image as the source domain and extract multi-granularity features from the source domain. Based on the image category information and the K-Means clustering algorithm, the prototype center c of the source domain is calculated. s ;

[0085] S303. Use ultra-high-definition images as the target domain and incorporate multi-granularity features of the target domain. The K-Means clustering algorithm is used to obtain the prototype centers c of the target domain. t ;

[0086] S304. Utilizing the multi-granularity features of the source domain image and the prototype center c of the source domain s And utilizing the multi-granularity features of the target domain image and the prototype center c of the target domain t In each domain, the domain cohesion loss is calculated.

[0087] S305, Multi-granularity features of joint source domain images and the prototype center c of the source domain s and multi-granular features of the target domain image and the prototype center c of the target domain t The inter-domain coupling loss is calculated between domains.

[0088] S306. Multi-granularity features of the source domain image Multi-granularity features of the target domain image The input is fed into the feature classifier to obtain the feature classification offset loss.

[0089] S307, Based on domain cohesion loss Inter-domain coupling loss and feature classification offset loss The total loss of the entire model optimization is calculated.

[0090] S308. Calculate the total loss. The gradient is used to backpropagate and optimize the parameters of the entire model.

[0091] S309. Determine whether the loss meets the convergence condition. If yes, obtain the trained ultra-high-definition image unsupervised adaptive recognition model and proceed to step S4. Otherwise, return to step S301.

[0092] It should be noted that, in order to extract multi-granularity features from the image, comprehensively perceive the fine-grained and feature-level features of the image, and obtain the image's depth encoding to facilitate the subsequent calculation of the prototype center and target loss function, the calculation formula for the multi-granularity features in step S301 of this embodiment is as follows:

[0093]

[0094]

[0095]

[0096] in, It is a multi-granularity feature set of batch images. It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. It is a multi-granularity feature extractor. and These are batch training datasets for source domain images and target domain images, respectively.

[0097] Meanwhile, in order to construct the prototype center of each category in the source and target domains, and to use this as a representative of the category and cluster feature distribution, laying the foundation for the calculation of cohesion loss and inter-domain coupling loss functions in subsequent domains, in step S302 of this embodiment, the prototype center c of the source domain... s The calculation formula is:

[0098]

[0099] The prototype center c of the target domain t The calculation formula is:

[0100]

[0101] Among them, c s This represents the prototype center of the constructed source domain. This represents the cluster centers obtained after performing K-Means clustering on the features of the source domain. Let represent the category center of each category in the source domain, e represent a weight hyperparameter used to balance the proportion of cluster centers and category centers in the construction of prototype centers, and c represent the category center. t This represents the prototype center of the target domain being constructed. Because the target domain lacks label information, only [the following is used]. To construct the prototype center, This represents the cluster centers obtained after K-Means clustering of the target domain features.

[0102] It should be noted that this embodiment employs in-domain cohesion loss, which uses the prototype centers in both the source and target domains as anchor points to shorten the distance between features of the same class (cluster) and the prototype centers. This makes features of the same class (cluster) more cohesive, improving the discriminative power of features in the feature space and establishing a certain degree of discriminative power foundation for subsequent feature classification and ultra-high-definition image recognition. Therefore, in step S304 of this embodiment, the in-domain cohesion loss... The expression is:

[0103]

[0104] in, The cohesion loss function in the domain, n s The total amount of source domain image data, n tThe total amount of image data in the target domain, i and j are indicator variables for iteratively calculating the loss function of the image data, and CE refers to the cross-entropy loss function. This represents the i′-th feature extracted from the source domain (conventional) image. This represents the extracted j′-th target domain (ultra-high definition) image feature. It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. Representative target domain features The index of the nearest cluster center.

[0105] It should be noted that this embodiment employs inter-domain coupling loss, which can use the cross-domain prototype center as an anchor point to shorten the distance between the feature clusters of the source domain (conventional image) and the feature clusters of the target domain (ultra-high-definition image) in the feature space, eliminate the domain offset between the two image domains, and improve the model's ability to supervise knowledge transfer. Therefore, in step S305 of this embodiment, the inter-domain coupling loss... The expression is:

[0106]

[0107] in, Represents the inter-domain coupling loss function. The cross-domain prototype center representing the feature cluster, Representative source domain features and target domain features The latest cross-domain prototype center c g The index.

[0108] Furthermore, in order to perform supervised learning of class labels on the conventional image features of the source domain, this embodiment employs cross-entropy loss constraint to encourage the class labels of the source domain to constrain the distribution of source domain features, thereby enhancing their discriminative power and enabling them to propagate to the target domain as inter-domain coupling deepens, thus improving the discriminative power of ultra-high-definition image features. Therefore, in step S306 of this embodiment, the feature classification shift loss... The expression is:

[0109]

[0110] in, Let f represent the feature classification offset loss, and let f represent the feature classifier that has been constructed. It is a function representation for classifying features.

[0111] It should be noted that this embodiment can combine the domain cohesion loss, inter-domain coupling loss, and feature classification offset loss function term through weighted fusion to achieve joint supervision of the objective function during model training. This promotes the model's ability to achieve domain adaptation and accurate ultra-high-definition image recognition and classification. Therefore, in step S307 of this embodiment, the expression for the total loss is:

[0112]

[0113] in, Represents the total loss function. α, β, and γ represent the domain cohesion loss function, the inter-domain coupling loss function, and the feature classification offset loss function, respectively, and represent the weight values ​​of the corresponding function terms.

[0114] Finally, in order to clarify the halting criteria for training the ultra-high-definition image unsupervised adaptive recognition model and to save the model when it reaches its optimal state and results, this embodiment specifies in step S309 that the total loss satisfies the convergence condition, meaning that the total loss of five consecutive batches satisfies the following:

[0115]

[0116] in, For the total loss of the ep-th batch, Let be the total loss for the (ep-1)th batch, and |·| be the absolute value function.

[0117] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-granularity sensing and prototype-driven ultra-high-definition image domain adaptive recognition method, characterized in that, Includes the following steps: Obtain regular image datasets and ultra-high-resolution image datasets across image domains, and preprocess the regular image datasets and ultra-high-resolution image datasets to obtain preprocessed batch training image datasets, including the following steps: Obtain a regular image dataset with accurate labels, which contains regular natural images of various scenes in production and daily life; Obtain an unlabeled ultra-high-resolution image dataset; Data cleaning was performed on the two image datasets mentioned above to remove erroneous and noisy images, resulting in a regular image dataset. and ultra-high-definition image datasets ; regular image datasets and ultra-high-definition image datasets The training data is divided and organized to obtain a batch training image dataset. ; Construct and initialize a multi-granularity feature extractor and a feature classifier; An unsupervised adaptive recognition model for ultra-high-definition images is trained based on a batch training image dataset, combining a multi-granularity feature extractor and a feature classifier. This process includes the following steps: Batch training image dataset Input to multi-granularity feature extractor In the process, multi-granularity features are extracted from each batch of images. ; Using a regular image as the source domain, and incorporating multi-granular features from the source domain... Based on the image category information and the K-Means clustering algorithm, the prototype centers of the source domain are calculated. ; Using ultra-high-resolution images as the target domain, and incorporating multi-granularity features of the target domain... The prototype centers of the target domain are obtained by using the K-Means clustering algorithm. ; Utilizing the multi-granularity features of source domain images and the prototype center of the source domain And utilizing the multi-granularity features of the target domain image and the prototype center of the target domain In each domain, the domain cohesion loss is calculated. ; Cohesion loss in the domain The expression is: in, The cohesion loss function in the domain represents the loss function. The total amount of source domain image data. Refers to the total amount of image data in the target domain. It is an indicator variable for iteratively calculating the loss function from image data. This represents the category label corresponding to the source domain feature. Cross-entropy loss function This indicates the extracted source domain image number 1. One characteristic, Indicates the extracted first Image features of the target domain It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. Representative target domain features The index of the nearest cluster center; Multi-granularity features of joint source domain images and the prototype center of the source domain and multi-granular features of the target domain image and the prototype center of the target domain The inter-domain coupling loss is calculated between domains. ; The inter-domain coupling loss The expression is: in, Represents the inter-domain coupling loss function. Refers to the calculation of information entropy. The cross-domain prototype center representing the feature cluster, Representative source domain features and target domain features The latest cross-domain prototype center Index; Multi-granular features of the source domain image Multi-granularity features of the target domain image The input is fed into the feature classifier to obtain the feature classification offset loss. ; According to the cohesion loss in the domain Inter-domain coupling loss and feature classification offset loss The total loss of the entire model optimization is calculated. ; Calculate total loss The gradient is used to backpropagate and optimize the parameters of the entire model. When the total loss meets the convergence condition, the trained ultra-high-definition image unsupervised adaptive recognition model is obtained. The ultra-high-definition image to be identified is input into an unsupervised adaptive recognition model based on ultra-high-definition images for classification, and the recognition result of the ultra-high-definition image is obtained.

2. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The construction and initialization of the multi-granularity feature extractor and feature classifier includes the following steps: Establish and initialize the Faster RCNN network framework with pre-trained parameters to obtain the object detection model. ; Establish and initialize the ResNet-18 network framework with pre-trained parameters to obtain the backbone network model. ; A single-layer convolutional neural network framework is established and randomly initialized to obtain a fine-grained attention network. A single-layer fully connected neural network framework was established to obtain a feature-granularity attention network. ; Combined target detection model Backbone network model Fine-grained attention networks and feature-granularity attention network A multi-granularity feature extractor was obtained. ; A multi-layer fully connected neural network framework is established and randomly initialized to obtain a feature classifier. .

3. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The calculation formula for the multi-granularity feature is as follows: in, It is a multi-granularity feature set of batch images. It is a multi-granularity feature set of the source domain image. It is a multi-granularity feature set of the target domain image. It is a multi-granularity feature extractor. and These are batch training datasets for source domain images and target domain images, respectively.

4. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The prototype center of the source domain The calculation formula is: The prototype center of the target domain The calculation formula is: in, This represents the prototype center of the constructed source domain. This represents the cluster centers obtained after performing K-Means clustering on the features of the source domain. This represents the category center of each category in the source domain. This represents a weighted hyperparameter used to balance the proportion of cluster centers and category centers in the construction of prototype centers. Indicates the prototype center of the target domain being constructed, using To construct the prototype center, This represents the cluster centers obtained after K-Means clustering of the target domain features.

5. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The feature classification offset loss The expression is: in, This represents the feature classification offset loss. This indicates that the feature classifier has been successfully constructed. It is a function representation for classifying features.

6. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The expression for the total loss is: in, Represents the total loss function. These represent the domain cohesion loss function, the inter-domain coupling loss function, and the feature classification offset loss function, respectively. These represent the weight values ​​of the corresponding function terms.

7. The ultra-high-definition image domain adaptive recognition method based on multi-granularity perception and prototype-driven methods according to claim 1, characterized in that, The total loss satisfying the convergence condition means that the total loss of 5 consecutive batches satisfies the following: in, For the first Total loss for each batch For the first Total loss for each batch It is an absolute value function.

Images