Plant disease small sample classification method and system based on feature space calibration

Abstract

The invention provides a plant disease small sample classification method and system based on feature space calibration, and belongs to the field of data processing. The invention aims to solve the problem that large-scale application of the deep learning algorithm in plant disease identification is limited due to the fact that a large number of data sets with annotations are needed during image classification by the existing deep learning algorithm, but part of plant diseases are low in morbidity and less in image data. And carrying out scoring analysis on each feature output layer of the Swinin-Transform V2 network after the full connection layer is removed by utilizing FAS scores, selecting the layer with the highest score as a feature vector to obtain a Swinin-Transform V2F6 network, and then carrying out plant disease feature space calibration. According to the method, the accuracy of small sample disease classification can be effectively improved, and compared with the need of using a large number of data sets with annotations, the recognition accuracy can be improved on the basis of a small number of data sets.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for classifying small samples of plant diseases based on feature space calibration. Background Art

[0002] Against the backdrop of the rapid development of hardware and information technology, deep learning has gradually become the mainstream method in the field of image recognition. In the conventional deep learning image classification process, a large number of images are required to input the model for training. If the model trained using the appropriate method can extract the features of the image well, it can show excellent performance in the testing phase. With the development of Internet of Things technology, more and more terminal devices are beginning to acquire massive amounts of images. Ordinary people can use high-definition mobile phone cameras to obtain real-life images of crop diseases around the world. Therefore, low-cost, general, automatic and accurate plant disease classification models are beginning to try to replace manual judgment of crop diseases.

[0003] At present, most of the best image classification results are obtained by deep learning algorithms based on convolutional neural networks (CNNs). A large amount of work has confirmed the feasibility of using deep learning technology for plant disease classification. However, one of the limitations of such algorithms is that they require a large number of annotated dataset images for training. However, in actual applications, due to the irregularity of the onset of some plant diseases in time and space, the probability of some diseases is low and the image data is small, which directly leads to the low recognition rate of such rare diseases, greatly limiting the large-scale application of deep learning technology in plant disease identification. Summary of the invention

[0004] The technical problems to be solved by the present invention are:

[0005] In order to solve the problem that existing deep learning algorithms require a large amount of annotated data sets for image classification, but the incidence of some plant diseases is low and the image data is scarce, which limits the large-scale application of deep learning algorithms in plant disease identification.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The present invention provides a plant disease small sample classification method based on feature space calibration, comprising the following steps:

[0008] S100, based on the Swin-Transformer V2 network, after removing the fully connected layer, input multiple feature vectors in sequence and run multiple plant disease data sets, and score each feature output layer in the Swin-Transformer V2 network by the feature layer evaluation index, wherein the feature evaluation index is the feature adaptation score FAS;

[0009] S200, selecting the output result FAS with the highest score in step S100 as the feature vector, deleting the Stage4 block before the last Patch Merging in the Swin-Transformer V2 network, and obtaining the Swin-Transformer V2 F6 network;

[0010] S300, calibrating the plant disease feature space according to the network structure Swin-Transformer V2 F6 obtained in step S100 and the Swin-Transformer V2 network designed in step S200.

[0011] Furthermore, in step S100, the plant disease dataset includes a PlantVillage dataset, a PlantDoc dataset, a Plant Real-World dataset and a Plant&Pest dataset, and FAS scoring is performed after seven feature vectors are input into the above four datasets.

[0012] Furthermore, the output result of the 6th layer after FAS scoring is used as the feature vector for small sample analysis.

[0013] Furthermore, in step S300, it specifically includes:

[0014] S310, the query set Q and the support set S are calculated by Swin-Transformer V2 F6 to obtain the feature vector Q features , S features , the training set P of the PlantVillage dataset is transformed into a feature vector set P by SwinTransformerV2F6 features , and as a benchmark, the eigenvector S features Perform calibration;

[0015] S320, for category i, the set of all feature vectors of PlantVillage P features The support set for this class in turn Find the L2 distance d i , sort all distances from small to large to get Take the first m vectors with the smallest distance after sorting Calculate its center of mass

[0016] Determine all the centroids c used for calibration i Whether the calculation is completed, if completed, proceed to the next step, if not completed, repeat this step until all are completed;

[0017] S330, the centroids of all categories calculated in step S320 are For class i, first calculate the class centroid c i and The L2 distance of and sorting it from small to large, we get Take the two distances with the smallest distance after sorting and The corresponding centroids are c a and c b , select a point T on the line connecting the two centroids; Move γ units along the direction that is collinear with itself and T to complete the support set The calibrated support set

[0018] Repeat the above steps until all support set features are completed. Calibration of

[0019] S340, calculate all calibrated support sets and the query set feature vector q features The distance d i , output the label corresponding to the query set q:

[0020] Furthermore, for 5-way-5-shot and 5-way-10-shot tasks, as the average of the support set features.

[0021] Furthermore, the location of point T is determined by the hyperparameter λ.

[0022] Furthermore, the calibrated support set is

[0023] A plant disease small sample classification system based on feature space calibration, the system has a program module corresponding to the above steps, and executes the steps in the above plant disease small sample classification method based on feature space calibration when running.

[0024] A computer-readable storage medium stores a computer program, wherein the computer program is configured to implement the steps of a plant disease small sample classification method based on feature space calibration when called by a processor.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The present invention discloses a method and system for classifying small samples of plant diseases based on feature space calibration. The method uses the FAS score to score and analyze each feature output layer of the Swin-Transformer V2 network, selects the layer with the highest score as the feature vector of the present invention, deletes the Stage 4 block of the previous layer of the last Patch Merging in the existing Swin-Transformer V2 network, and then calibrates the feature space of plant diseases. By comparing and analyzing with the existing algorithm, it can be seen that the feature extraction network structure Swin-Transformer V2 Compared with the existing network structure, F6 does not require training parameters. This structure can be directly obtained through a reasonable FAS. The performance of small sample disease classification can be significantly improved, especially in 5-way-1-shot, 5-way-5-shot and 5-way-10-shot tasks, the accuracy can be effectively improved. Compared with the need to use a large number of annotated data sets, this method can improve the recognition accuracy based on a small number of data sets, especially when each sample has only 10 pictures (5W10S), the performance on PlantVillage and FSPlantPest can reach the performance level of conventional supervised classification tasks, which is an important discovery for small sample tasks of plant diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a score diagram of the FAS indicator at each layer F1-F7 in the embodiment of the present invention;

[0028] Figure 2 This is a network structure diagram of Swin-Transformer V2 F6 in an embodiment of the present invention;

[0029] Figure 3 This is a comparison chart of the classification accuracy of Swin-Transformer V2 F6 and Swin-Transformer V2 F7 in 5-way-1-shot, 5-way-5-shot and 5-way-10-shot tasks respectively in an embodiment of the present invention;

[0030] Figure 4 It is a flow chart of a small sample classification method of plant diseases based on feature space calibration in an embodiment of the present invention;

[0031] Figure 5 1 is a diagram showing the implementation process of a small sample classification method for plant diseases based on feature space calibration in an embodiment of the present invention;

[0032] Figure 6This is a performance comparison chart of the method in the embodiment of the present invention and the existing algorithm on the PlantVillage data set;

[0033] Figure 7 This is a performance comparison chart of the method in the embodiment of the present invention and the existing algorithm on the PlantDoc dataset;

[0034] Figure 8 This is a performance comparison chart of the method in the embodiment of the present invention and the existing algorithm on the Plant Real-World dataset;

[0035] Fig. 9 This is a performance comparison chart of the method in the embodiment of the present invention and the existing algorithm on the Plant&Pest dataset. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0037] Specific implementation plan 1: Combine Figures 1 to 5 As shown, the present invention provides a small sample classification method for plant diseases based on feature space calibration, comprising the following steps:

[0038] S100. Based on the Swin-Transformer V2 network, each feature output layer in the Swin-Transformer V2 network is scored by the feature layer evaluation index. The feature evaluation index is the feature adaptation score FAS. In order to ensure the accuracy of the feature adaptation score FAS, currently public and common data sets including the PlantVillage data set, the PlantDoc data set, the Plant Real-World data set and the Plant&Pest data set are selected for analysis. The analysis results are as follows: Figure 1 As shown,

[0039] We assume that a network structure has a total of L layers, and we use φ l (x) represents the feature representation of the input image x in the lth layer of the network. The jth sample xij of category i in the training set is sent into the network and after L layers of forward inference, the image x is obtained. ij After each layer of features, it is flattened to obtain a one-dimensional feature vector φ l (x ij ), where l = 1, 2, ..., L;

[0040] If we use the ImageNet pre-trained model to directly calculate the feature vector, in order to adapt to different tasks, we need to remove the last fully connected layer of the network and use the feature vector φ of the penultimate layer, which is the L-1 layer L-1 (x) to compare with the prototype; The first goal of the present invention is to explore the L-1 layer feature vector φ that has always been considered to be most suitable for comparison with the prototype in the FSL task. L-1 (x) Whether it is also suitable for small sample tasks of crops;

[0041] To evaluate the feature vector φ of the lth layer l (x) Feature Adaptation Score (FAS) in the small sample task of plant diseases l , we first define three parameters, namely, intra-class variance Between-class variance and the average inter-class distance In order to derive these three parameters, we first need to calculate the mean μ of the feature vector of the sample of category i at layer l i,l And the feature vector mean μl of all categories at layer l is expressed as follows:

[0042]

[0043] Among them, N represents the total number of categories; K represents the total number of samples in each category;

[0044] Based on the above discussion, we derive expressions for three key parameters:

[0045]

[0046] Assume that the eigenvector φ l (x) is subject to a multidimensional normal distribution, then μ i,l and μ l It also obeys the multidimensional normal distribution, so σ2 obeys the non-central chi-square distribution; since the dimension directly affects the mean and variance of the non-central chi-square distribution, if you want to compare, you need to transform the eigenvector φ of each layer l (x) Scale it according to the dimension so that the feature vectors of all layers are compared on the same scale; finally, after dimensional scaling, we can define FAS as follows: l :

[0047]

[0048] The ideal feature vector φ used to calculate the distance to the prototype l(x), first of all, it should have low intra-class variance, high inter-class variance and high inter-class distance, so that the model can have good consistency for samples in the same category and high discrimination for samples in different categories; the more suitable the feature vector φ for the small sample task of plant diseases is, the better the model is. l The higher the FAS score corresponding to (x), the higher the value. Formula (6) can meet this requirement.

[0049] We use Swin-Transformer V2-T as the feature extractor for the small sample task of crop diseases. According to the official source code, after removing the fully connected layer, Swin-Transformer V2-T will output 7 features in sequence during the inference process. We use the pre-trained weights of ImageNet-1K and the backbone network of the Swin-TransformerV2-T structure to run the datasets PlantVillage, FSPlantPest, PlantDoc, and PlantReal-World in sequence, and obtain 7 sets of feature vectors φ for the four crop disease datasets. 1～7 (x), the corresponding FAS score is Figure 1 As shown;

[0050] from Figure 1 As can be seen in the table, the 6th layer feature vector of Swin-Transformer V2-T far exceeds the scores of all other layers, indicating that the 6th layer feature vector is more suitable for the small sample task of crop diseases than the feature vectors of other layers. At this time, we use the 6th layer feature vector F6 and the most commonly used 7th layer feature vector F7 to continue the comparative experiment on different data sets to test the accuracy of the small sample task. The results are as follows Figure 3 As shown;

[0051] From the experimental results, we can see that the highest-scoring F6 has achieved a significant improvement in the accuracy of small sample tasks compared to the commonly used F7. Therefore, whether through the comparison of the test set accuracy of FAS or small sample tasks, we can determine that F6 is more suitable for plant disease small sample tasks than F7. This is the first important discovery in the field of small samples. Based on the above experimental results, we can obtain the most suitable feature extraction network structure Swin-Transformer V2 F6 for plant disease small sample tasks, such as Figure 2 As shown;

[0052] Compared with the new structures proposed in previous small sample tasks, the biggest feature of the Swin-Transformer V2 F6 structure is that it has no training parameters and is a structure directly found through a reasonable FAS; this structure is directly inspired by Swin-Transformer V2, uses the official pre-training weights, and when each sample has only 10 images (5W10S), the performance on PlantVillage and FSPlantPest can reach the performance level of conventional supervised classification tasks, which is an important discovery for small sample tasks of plant diseases;

[0053] S200, combined Figure 2 As shown, unlike the existing Swin-Transformer V2 network, the present invention deletes the Stage4 block of the previous layer of the last Patch Merging in the Swin-Transformer V2 network;

[0054] It should be noted that Swin Transformer is an innovative neural network architecture designed for visual tasks. It extracts and utilizes image features by subdividing the input image into small non-overlapping blocks (called "tokens") and performing a series of complex transformations on these blocks.

[0055] Liu, Z., et al. Swin transformer: Hierarchical vision transformer using shifted windows. in Proceedings of the IEEE / CVF international conference on computer vision. 2021.

[0056] At this point, the main components of Swin Transformer and their functions are as follows:

[0057] (1) Image block segmentation and feature embedding: Swin Transformer first divides the input image into non-overlapping blocks of fixed size, which are regarded as independent "tokens";

[0058] Each image block is processed using a 4×4 convolution kernel, generating a feature dimension of 48 (4×4×3); these features are then transformed into a higher-dimensional feature space, denoted as C, through a linear embedding layer;

[0059] (2) Structure and function of Swin Transformer Block: The first stage of the network consists of two consecutive

[0060] Swin Transformer Block, each block processes the token while maintaining the total number of tokens (H / 4×W / 4);

[0061] The first Swin Transformer Block starts with a normalization layer (Layer Normalization, LN), followed by a multi-head self-attention mechanism (Multi-head Self-Attention, MSA), followed by another layer of LN and two layers of multi-layer perceptrons (MLP) with GELU nonlinear functions;

[0062] In the second Swin Transformer Block, Shifted Window MSA (SW-MSA) is used instead of MSA, and the other parts remain unchanged. This design optimizes context capture by adjusting the window position. The linear embedding layer and two consecutive Swin Transformer Blocks are marked as "Step 1";

[0063] (3) Hierarchical structure and feature integration: Swin Transformer adopts a hierarchical structure and gradually reduces the number of tokens through the "patchmerging" layer; in this process, each group of 2×2 adjacent patches is merged, and a linear transformation is applied to the merged 4C-dimensional features;

[0064] This operation effectively reduces the number of tokens by a quarter and doubles the feature dimension to 2C, thereby achieving a 2x reduction in spatial resolution. The following Swin Transformer blocks further refine these features, maintaining a resolution of H / 8×W / 8.

[0065] “Patch merging” and the subsequent two consecutive Swin Transformer Blocks are marked as “Step 2”. This process is repeated twice, marked as “Stage 3” and “Stage 4”, and the final output resolution reaches H / 16×W / 16 and H / 32×W / 32;

[0066] Swin Transformer V2 is a new visual Transformer architecture that has made several key improvements to the original Swin Transformer, aiming to improve the model's processing power and window resolution scalability; these improvements make Swin Transformer V2 more efficient and flexible in handling a variety of visual tasks, especially for applications that require advanced feature integration capabilities, such as image smoothing, object detection, and semantic segmentation;

[0067] At this time, the core improvements of Swin Transformer V2 (compared with V1) include: (1) Post-normalization layer: In the V2 version, the post-normalization layer is introduced to replace the pre-normalization layer in the previous version; this new configuration helps the model to be more stable when expanding its capacity and simplifies the training process; (2) Scaled Cosine Attention Mechanism: The scaled cosine attention mechanism introduced in the V2 version replaces the traditional dot-product attention; scaled cosine attention improves the flexibility and efficiency of the model in processing features of different scales by adjusting the scaling factor of the attention mechanism; (3) Log-Spaced Continuous Relative Position Bias: V2 adopts a new position bias method, which expresses the relative position relationship in a logarithmic spaced continuous form. This method has better transferability between different window resolutions and improves the accuracy of the model when processing large-size inputs; these innovative features make Swin Transformer While maintaining local information, V2 can also effectively capture global context, which is particularly critical for image processing tasks. In image smoothing applications, these features of Swin Transformer V2 can be used to accurately process high-frequency and low-frequency information of images and optimize the overall visual effect of the image.

[0068] S300, combined Figure 4 and Figure 5 As shown, the plant disease feature space calibration is performed according to the network structure Swin-Transformer V2F6 obtained in step S100 and the Swin-Transformer V2 network designed in step S200, specifically including:

[0069] S310, the query set Q and the support set S are calculated by Swin-Transformer V2 F6 to obtain the feature vector Q features , S features, the training set P of the PlantVillage dataset is transformed into a feature vector set P by SwinTransformerV2F6 features , and as a benchmark, the eigenvector S features Perform calibration;

[0070] S320, for category i, the set of all feature vectors of PlantVillage P features The support set for this class in turn Find the L2 distance d i , sort all distances from small to large to get Take the first m vectors with the smallest distance after sorting Calculate its center of mass

[0071] Determine the centroids c of all the calibration i Whether the calculation is completed, if completed, proceed to the next step, if not completed, repeat this step until all are completed;

[0072] S330, we will proceed based on the centroid of all categories calculated by this method For class i, we first calculate the class centroid c i and The L2 distance of Take the two distances with the smallest distance after sorting The corresponding centroid c a 、c b , select a point T on the line connecting the two centroids, and its position is determined by the hyperparameter λ; at this time, Move γ units along the direction that is collinear with itself and T to complete the support set The calibrated support set

[0073] Repeat the above steps until all support set features are completed. The calibration is completed;

[0074] S340, finally calculate all calibrated support sets and the query set feature vector q features The distance d i , we can output the label corresponding to the query set q: For 5-way-5-shot and 5-way-10-shot tasks, just change The average value of the support set features can be used.

[0075] Specific implementation scheme 2: The present invention provides a plant disease small sample classification system based on feature space calibration, which has a program module corresponding to the above steps and executes the steps in the above plant disease small sample classification method based on feature space calibration during operation.

[0076] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0077] Specific implementation scheme three: The present invention provides a computer-readable storage medium, which stores a computer program, and the computer program is configured to implement the steps of a small sample classification method of plant diseases based on feature space calibration when called by a processor.

[0078] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0079] Comparative experiment

[0080] Hardware: The experiments were conducted on a computer equipped with an Intel i7 processor and 16GB RAM.

[0081] Software environment: Use Python 3.9, relying on NumPy and PyTorch libraries.

[0082] Dataset: Public image classification datasets used include PlantVillage dataset, PlantDoc dataset, Plant Real-World dataset, and Plant&Pest dataset.

[0083] Experimental process: After the software environment has been installed, follow the process Figure 2 By executing the algorithm, the effects mentioned in the invention can be achieved.

[0084] Application scenarios: This algorithm is suitable for plant disease classification and computer image processing scenarios.

[0085] The present invention uses Swin-Transformer V2 F6 as a fixed feature extractor, and uses the feature space calibration algorithm to calibrate the feature space of the features extracted from the structure. At the same time, we compare the performance results of Swin-Transformer V2F6 and the feature space calibration algorithm with the latest methods in the field of Plant Disease Few Shot (PDFS). The comparison evaluation indicators are standard evaluation indicators in small sample learning: classification accuracy of 5-way-1-shot, 5-way-5-shot and 5-way-10-shot. We used various methods in the PDFS tasks of the past three years for comparison, and the comparison results of various methods on four data sets are shown in Figures 6 to 9 From the results, it is not difficult to see that the classification accuracy of the Swin-Transformer V2F6 feature space calibration structure on the PDFS task comprehensively exceeds the new methods in recent years.

[0086] exist Figures 6 to 9 The method used is from:

[0087] The source of Li er al. (2021) is Y. Li and J. Yang, "Meta-learning baselines and database for few-shot classification in agriculture," Comput.

[0088] The source of Yang et al. (2022) is J. Yang, Y. Shuang Yang, Y. Li, S. Xiao, and S. Ercis, li, "Image information contribution evaluation for plant diseases classification via inter-class similarity," Sustainability, 2022. [Online]. Available: https: / / api.semanticscholar.org / CorpusID:252060364.

[0089] The source of Huang et al. (2022) is Y. Huang, F. Chang, Y. Tao, Y. Zhao, L. Ma, and H. Su, “Fewshot learning based on attn-cutmix and task-adaptive transformer for therecognition of cotton growth state," Comput. Electron. Agric., vol. 202, p. 107406, 2022. [Online]. Available: https: / / api.semanticscholar.org / Corpu sID: 252708627.

[0090] The source of Zhang et al. (2021) is D.Zhang, F.Pan, Q.Diao,

[0091] The source of Rezaei et al. (2024) is M. Rezaei, D. Diepeveen, H. Laga, MGK Jones, and F. Sohel, “Plant disease recognition in a low data scenario using few-shot learning,” Comput.

[0092] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A small sample classification method for plant diseases based on feature space calibration, characterized in that: The following steps are involved: S100, based on the Swin-Transformer V2 network, after removing the fully connected layer, input multiple feature vectors in sequence and run multiple plant disease data sets, and score each feature output layer in the Swin-Transformer V2 network by the feature layer evaluation index, wherein the feature evaluation index is the feature adaptation score FAS; S200, select the output result FAS with the highest score in step S100 as the feature vector, delete the Stage4 block of the previous layer of the last Patch Merging in the Swin-TransformerV2 network, and obtain the Swin-Transformer V2F6 network; S300, calibrating the plant disease feature space according to the network structure Swin-Transformer V2 F6 obtained in step S100 and the Swin-Transformer V2 network designed in step S200.

2. The method for plant disease small sample classification based on feature space calibration according to claim 1, characterized in that: In step S100, the plant disease datasets include PlantVillage dataset, PlantDoc dataset, PlantReal-World dataset and Plant&Pest dataset, and FAS scoring is performed after seven feature vectors are input into the above four datasets.

3. The method for plant disease small sample classification based on feature space calibration according to claim 2, characterized in that: The output of the 6th layer after FAS scoring is used as the feature vector for small sample analysis.

4. The method for plant disease small sample classification based on feature space calibration according to claim 3, characterized in that: In step S300, it specifically includes: S310, the query set Q and the support set S are calculated by Swin-Transformer V2 F6 to obtain the feature vector Q features , S features , the training set P of the PlantVillage dataset is transformed into a feature vector set P by SwinTransformerV2F6 features , and as a benchmark, the eigenvector S features Perform calibration; S320, for category i, the set of all feature vectors of PlantVillage P features The support set for this class in turn Find the L2 distance d i , sort all distances from small to large to get Take the first m vectors with the smallest distance after sorting Calculate its center of mass Determine all the centroids c used for calibration i Whether the calculation is completed, if completed, proceed to the next step, if not completed, repeat this step until all are completed; S330, the centroids of all categories calculated in step S320 are For class i, first calculate the class centroid c i and The L2 distance of and sorting it from small to large, we get Take the two distances with the smallest distance after sorting and The corresponding centroids are c a and c b , select a point T on the line connecting the two centroids; Move γ units along the direction that is collinear with itself and T to complete the support set The calibrated support set Repeat the above steps until all support set features are completed. Calibration of S340, calculate all calibrated support sets and the query set feature vector q features The distance d i , output the label corresponding to the query set q:

5. The method for plant disease small sample classification based on feature space calibration according to claim 4, characterized in that: For 5-way-5-shot and 5-way-10-shot tasks, as the average of the support set features.

6. The method for plant disease small sample classification based on feature space calibration according to claim 5, characterized in that: The location of point T is determined by the hyperparameter λ.

7. The method for plant disease small sample classification based on feature space calibration according to claim 6, characterized in that: The calibrated support set is 8. A plant disease small sample classification system based on feature space calibration, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 7, and executes the steps in the above-mentioned plant disease small sample classification method based on feature space calibration when running.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the plant disease small sample classification method based on feature space calibration according to any one of claims 1 to 7 when called by a processor.

Images