Mass multi-mode data retrieval method oriented to Internet of Vehicles

Abstract

The invention belongs to the technical field of Internet of Vehicles data management, and particularly relates to an Internet of Vehicles-oriented massive multimode data retrieval method, which comprises the following steps of: respectively extracting image features and text features from a license plate image Xr and a description text Tm of target data through a visual encoder and a text encoder of a CLIP model; carrying out dimension reduction on the image features through a PCA algorithm; splicing the obtained text features and the image features after dimension reduction to obtain image text joint features; compressing the image text joint features through an MM-VAE model to obtain compressed features; and calculating the cosine similarity between the query feature and each target data compression feature, and finding out the most matched target data. According to the retrieval method disclosed by the invention, the multi-modal data features can be efficiently compressed and fused in a unified potential space, so that the retrieval accuracy and efficiency are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of vehicle networking data management, and in particular relates to a massive multi-modal data retrieval method for vehicle networking. Background Art

[0002] In recent years, with the increasing adoption of 5G technology and various sensor devices, connected vehicle (IoV) technology has matured. A large number of new energy connected vehicles (NEVs) generate massive amounts of multimodal data during their journeys. This multimodal data comes in a wide variety of forms, including sensor signals (such as speed, temperature, and pressure), image and video data (such as road conditions captured by onboard cameras), text data (such as vehicle status reports and driving logs), and spatiotemporal trajectory data (such as GPS location information). This multimodal data provides a valuable resource and foundation for advancing the digitalization, automation, and intelligence of vehicles. However, this data varies not only in type but also in volume and update frequency. This diversity and complexity places high demands on data storage and processing. To address the need to manage massive amounts of multimodal data, a new generation of cloud-native databases has emerged, equipped with the ability to integrate and process multimodal data. For example, Alibaba Cloud's Lindorm multimodal database supports multiple data models, including time series data, wide table data, and object storage data, providing a unified data storage and management platform. However, although cloud-native databases have shown great potential in multi-modal data processing, they still face many technical challenges when applied to Internet of Vehicles data management, mainly in terms of data storage and retrieval.

[0003] The storage of massive time-series data faces the challenge of exponential growth over time and the inability of existing storage solutions to cope in the long term. While solutions like NoSQL databases have alleviated data access pressure to some extent, solutions for massive time-series data are still immature and computer system resource utilization is low. Regarding multimodal data retrieval, the Internet of Vehicles (IoV) environment contains a vast amount of structured, semi-structured, and unstructured data, requiring fused retrieval to meet diverse information needs. Traditional database systems struggle to efficiently handle fused retrieval of multimodal data.

[0004] Therefore, how to efficiently store, manage and utilize this data has become an urgent problem to be solved in the field of Internet of Vehicles. Summary of the Invention

[0005] In order to address the deficiencies in massive time-series data compression and multimodal data retrieval in smart cars and the Internet of Vehicles, and to improve the efficiency of cross-modal data retrieval of license plate images and description texts, the present invention provides a massive multimodal data retrieval method for the Internet of Vehicles, so as to achieve accurate and fast retrieval of multimodal data in the Internet of Vehicles.

[0006] To solve the above technical problems, the present invention adopts a technical solution: a massive multimodal data retrieval method for the Internet of Vehicles, which is implemented based on the PCA-MM-VAE model and includes the following steps:

[0007] Step 1: Through the visual encoder and text encoder of the CLIP model, the license plate image X of the target data is obtained. r , description text T m Extract image features and text features respectively;

[0008] Step 2: Use the PCA algorithm to reduce the dimension of the image features obtained in step 1;

[0009] Step 3: Concatenate the text features obtained in step 1 and the image features after dimensionality reduction obtained in step 2 to obtain the joint image-text features;

[0010] Step 4: Compress the image-text joint features obtained in step 3 through the MM-VAE model to obtain compressed features;

[0011] Step 5: Calculate the cosine similarity between the query feature and the compressed features of each target data to find the best matching target data.

[0012] In step 1, the dimension of the image features extracted by the CLIP model is 768, and the dimension of the text features is 512.

[0013] In step 1, the CLIP model is trained using the AdamW optimizer, and the learning rate is set to 1E-6 during parameter optimization.

[0014] In step 1, the loss function of the CLIP model is:

[0015]

[0016] Where N represents the target data volume, z i ,t i Represent the image embedding vector and text embedding vector of the i-th target data, t j represents the text embedding vector of the j-th target data, τ represents the temperature scaling factor, and CS represents the cosine similarity.

[0017] In step 2, the image features obtained in step 1 are reduced in dimension to 512 dimensions using the PCA algorithm.

[0018] In step 4, the MM-VAE model compresses the image-text joint features through a variational autoencoder. The compression process is expressed as:

[0019]

[0020] Among them, Z compressed represents the compression feature, d 64 Represents the compressed feature dimension, MM-VAE encoder represents the encoder in MM-VAE, Z joint Represents the joint features of image and text.

[0021] In step 4, when the image-text joint features obtained in step 3 are compressed using the MM-VAE model, the dimension of the obtained compressed features is 64.

[0022] In step 5, the calculation formula of cosine similarity is:

[0023]

[0024] Among them, CS represents cosine similarity, Z q represents the query feature, Z compressed Indicates the compression characteristics of the target data.

[0025] In step 5, the query features include image and text features.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention provides a massive multimodal data retrieval method for the Internet of Vehicles. By combining the high-dimensional image and text features extracted by the CLIP model, and using PCA to reduce the dimensionality of the image features, the joint features of the image and text are compressed through the MM-VAE model. The cosine similarity between the query features and each target data is calculated to find the most matching target data, thereby realizing the query of multimodal data. The present invention improves the similarity measurement accuracy between different modal data through common subspace and deep learning technology, and makes significant progress in the accuracy and efficiency of cross-modal retrieval, successfully reducing the storage space and computing resources occupied, and greatly improving the speed of the retrieval process. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic diagram of a flow chart of a method for retrieving massive multimodal data for Internet of Vehicles provided by an embodiment of the present invention;

[0029] Figure 2 Schematic diagram of training loss and validation loss curves obtained in an embodiment of the present invention;

[0030] Figure 3 1 and 2 are curves of the recall rate (Recall) obtained in the embodiment of the present invention on different data, where (a) represents Recall@1, (b) represents Recall@10, and (c) represents Recall@50. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] Example 1

[0033] The first embodiment of the present invention provides a massive multimodal data retrieval method for the Internet of Vehicles (IoV), aiming to complete the task of efficiently retrieving cross-modal data based on vehicle images and vehicle description text. Specifically, given a multimodal query task (i.e., a vehicle image and its description text), the present invention uses the CLIP model to extract feature vectors of the image and text data respectively, then uses the PCA principal component analysis method to reduce the dimensionality of the extracted image, and finally obtains the compressed features of the image-text fusion through MM-VAE model training, and implements multimodal image retrieval based on this compressed feature.

[0034] Specifically, if Figure 1 As shown, this embodiment includes the following steps:

[0035] Step 1: Through the visual encoder and text encoder of the CLIP model, the license plate image X of the target data is obtained. r , description text T m Image features and text features are extracted separately.

[0036] The CLIP model used in step 1 is an image representation model learned through natural language supervision. It combines an image encoder and a text encoder and is pre-trained on a large amount of image and text paired data. Training is performed by maximizing the cosine similarity in the embedding space of images and text. This makes the CLIP model capable of zero-shot tasks, meaning it can be directly applied to various visual tasks without training on a specific dataset.

[0037] The CLIP model used in this example is a pre-trained model. The pre-training process uses a contrastive learning method, in which image and text pairs are used as training samples. By jointly training the image and text encoders, a multimodal embedding space is learned, which enables information from different modalities (image and text) to be effectively aligned in the same space. Assuming that the target data is N groups, that is, there are N image-text pairs, which can be expressed as:

[0038]

[0039] Where F represents the target data set, (X r,T m ) i represents the i-th group of target data, where X r ,T m They are the license plate image and description text of the target data respectively; through the visual encoder and text encoder of the CLIP model, from the license plate image X r , description text T m The specific process of extracting image features and text features can be expressed as:

[0040]

[0041] Among them, x r Represents the extracted image features, CLIP vision represents the visual encoder in the CLIP model, d image Represents the dimension of image features, t m Represents the extracted text features, CLIP text represents the text encoder in the CLIP model, d text Dimensions representing text features.

[0042] The goal of the CLIP model is to maximize the cosine similarity of correct image-text pairs and minimize the cosine similarity of incorrect pairs by optimizing the embedding space of images and text. Specifically, the loss function used by the CLIP model in this embodiment can be expressed as:

[0043]

[0044] Where N represents the target data volume, z i and t i Represent the embedding vectors of image and text respectively, t j represents the text embedding vector of the j-th target data, τ represents the temperature scaling factor, CS represents the function for calculating cosine similarity, CS(z i ,t i ) represents the cosine similarity between the embedding vectors of the image and the text.

[0045] In this embodiment, the contrastive learning framework of the CILP model is used to optimize the extracted image features and text features by maximizing the cosine similarity of image and text pairs, making the multimodal representation more consistent.

[0046] Specifically, in this embodiment, the dimension of the image features extracted by the CLIP model is 768, and the dimension of the text features is 512.

[0047] Specifically, in step 1 of this embodiment, the CLIP model is trained using the AdamW optimizer, and the learning rate is set to 1E-6 during parameter optimization.

[0048] Step 2: Use the PCA algorithm to reduce the dimensionality of the image features obtained in step 1.

[0049] Specifically, in step 2, the image features obtained in step 1 are reduced in dimension to 512 dimensions using the PCA algorithm.

[0050] Since the dimensions of image features and text features are high, storage and computation may bring a heavy burden. In addition, there is a situation where the dimensions of visual features and text features are not uniform. In this embodiment, the dimensions are 768 and 512 respectively. Therefore, PCA dimensionality reduction is performed on the image features. By reducing the image features from 768 dimensions to 512 dimensions through PCA, the storage space requirement and computational complexity of the data can be reduced. This process can be expressed as:

[0051]

[0052] where d reduced represents the feature dimension after dimensionality reduction, x r reduced represents the image features after dimensionality reduction, and PCA represents the PCA algorithm.

[0053] Specifically, in step 2, in order to avoid the influence of different feature scales between different images on the PCA algorithm, the image features need to be standardized first. The standardization step is to process each image feature value of the image so that the mean of all features is zero and the standard deviation is 1. The standardization formula is:

[0054]

[0055] Where μ and σ are the mean and standard deviation of the image features, respectively. X0 represents the matrix formed by N original images, and X1 represents the normalized image.

[0056] Secondly, the covariance matrix ∑ is calculated to represent the relationship between features. The calculation formula is:

[0057]

[0058] Solve for the eigenvalue λ of the covariance matrix ∑ i and the eigenvector ν i :

[0059] ∑v i =λ i v i (8)

[0060] The larger the eigenvalue, the greater the data variance in the direction of the eigenvector, which means that the direction contains more data variability and information. Based on the size of the eigenvalue, select the eigenvectors with the largest eigenvalues. These eigenvectors are the new coordinate axis directions, representing the main information (or variance) of the data. Usually, the first few principal components can capture most of the information in the image. Project the original data onto the first k principal components to achieve dimensionality reduction:

[0061] Y=X1W k (9)

[0062] Where Y is the data after dimensionality reduction, which is represented as an N×k matrix, k represents the number of principal components, W k It is a p×k matrix (p is the product of the image size and the channel).

[0063] The PCA algorithm achieves the effect of image feature dimensionality reduction through the above process, which can effectively reduce information loss and reduce data dimension.

[0064] Step 3: Concatenate the text features obtained in step 1 and the reduced-dimensional image features obtained in step 2 to obtain the joint image-text features.

[0065] In this embodiment, the image feature x after dimension reduction in step 2 is r reduced and text features t m Splice and get an image-text joint feature Z joint ,Right now:

[0066]

[0067] Among them, d joint =d reduced +d text is the dimension of the joint feature after concatenation. For example, if the visual feature is reduced to 512 dimensions and the text feature is 512 dimensions, then d joint =1024.

[0068] Step 4: Compress the image-text joint features obtained in step 3 through the MM-VAE model to obtain compressed features.

[0069] In step 4, the MM-VAE model maps the joint features to the latent space through the framework of the variational autoencoder to learn a compact latent representation, that is, the concatenated joint features (1024 dimensions) are further compressed to a lower dimension (e.g., 64 dimensions). This latent representation can retain the key information between the image and text while reducing redundancy.

[0070] Specifically, the variational autoencoder of the MM-VAE model receives the high-dimensional joint feature Z jointInput, processed by the neural network, output latent space mean μ joint and variance σ joint 2 Parameters, generate the distribution of latent variables, usually assuming that the distribution is Gaussian distribution, the generated latent variable distribution is:

[0071]

[0072] Among them, q θ (z│Z joint ) represents the joint feature Z joint The probability distribution in the latent space, N represents a Gaussian distribution, and z represents a latent variable.

[0073] In this embodiment, the MM-VAE model uses a reparameterization technique to sample the latent variable z from the distribution of the latent space. According to the reparameterization technique, the latent variable is represented as:

[0074] z=μ joint +σ joint ·∈,∈~N(0,I); (12)

[0075] where ∈ is the noise sampled from the standard normal distribution N(0,I).

[0076] Through the regularization of the latent space, the MM-VAE model further compresses these fused features and learns the latent space to obtain low-dimensional fused features, whose compression expression can be expressed as:

[0077]

[0078] Among them, d 64 =64 represents the feature dimension after compression, MM-VAE encoder Represents the encoder part in MM-VAE. Z compressed Indicates the compression feature. In this embodiment, the compressed Z compressed The feature dimension is 64, which greatly reduces the amount of data for subsequent storage and retrieval.

[0079] Step 5: Calculate the cosine similarity between the query feature and the compressed features of each target data to find the best matching target data. Specifically, the cosine similarity is used for sorting to find the best matching target data. In this embodiment, the target data with the largest cosine similarity is selected as the best matching data.

[0080] Specifically, in step 5, the calculation formula of cosine similarity is:

[0081]

[0082] Among them, CS represents cosine similarity, Z q represents the query feature, Z compressed Indicates the compression characteristics of the target data.

[0083] Specifically, in step 5, the query features include two features: image and text.

[0084] In step 5 of this embodiment, the DQU-CIR model is used to complete cross-modal query for text uniformity. In this process, image retrieval is performed by reading the 64-dimensional joint feature representation of the image and text and calculating the cosine similarity between the query features and the target image features.

[0085] In this example, the public datasets CCPD and Shoes datasets were selected to evaluate the data retrieval method of the present invention. The CLIP parameters were fine-tuned with a learning rate of 1E-6, and other parameters were optimized at 1E-4 to achieve effective convergence. The batch size was fixed at 16, and the temperature factor τ in the loss function of the CLIP model was set to 0.1. All experiments were performed using Pytorch on a server equipped with a single A100-40G GPU.

[0086] The evaluation follows the standard evaluation protocol of the dataset, using training loss (train loss) and validation loss (valloss) to reflect the learning ability and generalization ability of the model. At the same time, Recall@K (R@K) and the training time per round are used as indicators for fair comparison, where Recall mainly focuses on R@1, R@10 and R@50. By comparing the Recall@K values of the original model on the original dataset (shoes) and the new dataset (CCPD), it is evaluated whether the new dataset is suitable for the multimodal retrieval model and maintains high performance. By comparing the Recall@K values of the new data on the original model (DQU-CIR) and the improved model (MCR) of the embodiment of the present invention, it is evaluated whether the improved model will introduce data distortion, thereby resulting in a decrease in accuracy.

[0087] Dataset: This embodiment of the present invention selects a dataset based on the public dataset CCPD[5] and the public dataset shoes[6,9] to evaluate the improved model based on DQU-CIR.

[0088] Implementation Details: The MCR model in this embodiment uses pre-trained CLIP (clip-vit-base-patch32 version) for feature extraction and is trained using the AdamW optimizer. For the CCPD and Shoes datasets, CLIP parameters are fine-tuned with a learning rate of 1E-6, and other parameters are optimized at a learning rate of 1E-4 to achieve effective convergence. The batch size is fixed at 16, and the temperature factor τ in Equation (10) is set to 0.1. All experiments are conducted using PyTorch on a server equipped with a single A100-40G GPU.

[0089] Evaluation: The model evaluation follows the standard evaluation protocol of the dataset, using training loss (train loss) and validation loss (val loss) to reflect the model's learning and generalization capabilities. At the same time, Recall@K (R@K) and the training time per round are used as fair comparison indicators, where Recall mainly focuses on R@1, R@10, and R@50. By comparing the Recall@K values of the original model on the original dataset (shoes) and the new dataset (CCPD), it is evaluated whether the new dataset is suitable for multimodal retrieval models and maintains high performance. By comparing the Recall@K values of the new data on the original model (DQU-CIR) and the new model (MCR), it is evaluated whether the improved model introduces data distortion, resulting in a decrease in accuracy.

[0090] Experimental Results: The CCPD dataset and the Shoes dataset used in the original model were applied to the original model (DQU-CIR), and the Recall@K of the original model was calculated to ensure that the CCPD dataset is suitable for the retrieval model. As shown in Table 1, the dataset created based on the CCPD dataset in this embodiment has a high degree of compatibility with the multimodal retrieval model, with an average Recall@K of 27.20% higher than that of the original model data. This data performs well in the multimodal retrieval model.

[0091] Table 1: Comparison of Recall@K (%) in the original model (DQU-CIR) for the CCPD dataset and the Shoes dataset

[0092]

[0093] This table compares the performance of the CCPD dataset and the Shoes dataset in the original DQU-CIR model. The data shows that the CCPD dataset has a higher recall rate than the Shoes dataset at different K values, with an average increase of 27.20%. The R@1 rate also increases by 42.01%, demonstrating superior matching results in retrieval tasks.

[0094] The CCPD dataset was applied to the improved compressed retrieval model (MCR) of this embodiment, and the Recall@K of the original model and the improved model for the same dataset were compared. The experimental results show that its Recall@K decreased slightly, with R@1 decreasing by 1.51%, R@10 decreasing by 0.79%, and the overall loss rate being 0.87% (see Table 2). Despite this, the decrease is small, indicating that the improved model of this embodiment can still maintain a high retrieval performance while reducing computational complexity. Compared with the improvement in computing time and efficiency (see Table 3), this is negligible, further verifying the potential of the MCR model of this embodiment to reduce training time and save computational overhead.

[0095] Table 2: Comparison of Recall@K (%) between the two models on the CCPD dataset

[0096]

[0097]

[0098] from Figure 2 From the changing trends of the training loss (Train Loss) and validation loss (Val Loss) in the training data, the training process of the model of this embodiment performs well. In the early stage of training, the training loss drops rapidly, indicating that the model can quickly learn and adapt to the training data. Although the validation loss fluctuates in the early stage, as the training progresses, the validation loss gradually decreases steadily, showing that the model not only performs well on the training set, but also maintains good generalization ability on the validation set. In the end, both the training loss and the validation loss tend to be stable, indicating that the model converges stably during the training process without overfitting. Overall, the training process of the model is healthy and has good generalization ability, and is expected to achieve ideal performance in practical applications. Both the training loss and the validation loss decrease and stabilize with the training process, indicating that the model has successfully converged, has good generalization ability, and has no obvious overfitting.

[0099] according to Figure 3 As shown in the Recall@1, Recall@10, and Recall@50 curves, the MCR model of this embodiment shows a significant performance improvement during the training process. In the early stages, all three indicators experienced fluctuations, but as training progressed, the performance of the model gradually stabilized and improved. Ultimately, Recall@1, Recall@10, and Recall@50 all reached high levels, indicating that the model of this embodiment can maintain high accuracy across different retrieval ranges. The gradual improvement of these three indicators reflects the stability of the model of this embodiment when processing different retrieval tasks, indicating that the model of this embodiment has strong adaptability and robustness in multimodal data processing and retrieval tasks.

[0100] Table 3: Comparison of training time between the original model and the improved model for the same number of CCPD datasets

[0101]

[0102] By comparing the training time of the original model and the improved model of this embodiment under the same data volume, it was found that the training time of the improved model of this embodiment was significantly reduced compared to the original model, reaching a reduction of 88.26%. This result shows that the improved model of this embodiment has achieved significant improvements in computational efficiency. Under the same data volume, it can significantly shorten the training time, improve the efficiency of the training process, and reduce computational overhead.

[0103] In summary, the present invention provides a massive multimodal data retrieval method for the Internet of Vehicles, which converts license plate images and description texts into shared 64-dimensional compressed features, and completes the extraction, fusion and cross-modal query of image and text features through an improved method. The present invention not only reduces storage requirements and computational complexity, but also improves the speed and accuracy of the retrieval process, especially when dealing with complex cross-modal retrieval tasks. Based on the fusion of PCA dimensionality reduction, MM-VAE compression and CLIP models, the retrieval method of the present invention can efficiently compress multimodal data features and fuse them in a unified latent space, significantly improving the accuracy and efficiency of retrieval. Experimental results show that the improved model has achieved excellent performance on multiple standard data sets, especially in terms of indicators such as Recall@1, Recall@10 and Recall@50, which proves the strong adaptability and robustness of the method in complex cross-modal tasks.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A massive multimodal data retrieval method for the Internet of Vehicles, characterized by: The following steps are involved: Step 1: Through the visual encoder and text encoder of the CLIP model, the license plate image X of the target data is obtained. r , description text T m Extract image features and text features respectively; Step 2: Use the PCA algorithm to reduce the dimension of the image features obtained in step 1; Step 3: Concatenate the text features obtained in step 1 and the image features after dimensionality reduction obtained in step 2 to obtain the joint image-text features; Step 4: Compress the image-text joint features obtained in step 3 through the MM-VAE model to obtain compressed features; Step 5: Calculate the cosine similarity between the query feature and the compressed features of each target data to find the best matching target data.

2. A method for retrieving massive multimodal data for Internet of Vehicles according to claim 1, characterized in that: In step 1, the dimension of the image features extracted by the CLIP model is 768, and the dimension of the text features is 512.

3. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 1, characterized in that: In step 1, the CLIP model is trained using the AdamW optimizer, and the learning rate is set to 1E-6 during parameter optimization.

4. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 1, characterized in that: In step 1, the loss function of the CLIP model is: Where N represents the target data volume, z i ,t i Represent the image embedding vector and text embedding vector of the i-th target data, t j represents the text embedding vector of the j-th target data, τ represents the temperature scaling factor, and CS represents the cosine similarity.

5. The method for retrieving massive multimodal data for Internet of Vehicles according to claim 2, characterized in that: In step 2, the image features obtained in step 1 are reduced in dimension to 512 dimensions using the PCA algorithm.

6. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 1, characterized in that: In step 4, the MM-VAE model compresses the image-text joint features through a variational autoencoder. The compression process is expressed as: Among them, Z compressed represents the compression feature, d 64 Represents the compressed feature dimension, MM-VAE encoder represents the encoder in MM-VAE, Z joint Represents the joint features of image and text.

7. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 2, characterized in that: In step 4, when the image-text joint features obtained in step 3 are compressed using the MM-VAE model, the dimension of the obtained compressed features is 64.

8. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 1, characterized in that: In step 5, the calculation formula of cosine similarity is: Among them, CS represents cosine similarity, Z q represents the query feature, Z compressed Indicates the compression characteristics of the target data.

9. The method for retrieving massive multimodal data for the Internet of Vehicles according to claim 1, characterized in that: In step 5, the query features include image and text features.

Images