Zero-shot picture classification method, system, device and medium based on collaborative learning
By using a collaborative learning approach, classifier 1 and classifier 2 are trained using an image encoder and a semantic encoder, which solves the problem of failing to effectively utilize complementary feature information in existing technologies and improves the accuracy of zero-shot image classification.
Patent Information
- Application Number
- CN202211486711.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-24
- Publication Date
- 2026-08-04
- Estimated Expiration
- 2042-11-24
AI Technical Summary
Existing zero-shot image classification methods fail to effectively utilize the complementary information of image features and semantic features during the training phase, resulting in insufficient classification accuracy.
A collaborative learning approach is adopted, which involves building an image encoder and a semantic encoder to extract image and semantic codes respectively. Classifier 1 and Classifier 2 are trained using two subsets of datasets. Through cross-validation and iterative optimization, the prediction results of the two are finally combined to improve classification accuracy.
By using collaborative training to integrate image features and semantic features, the ability to identify unseen classes was improved, thus increasing the accuracy of zero-shot image classification.
Smart Images
Figure CN115908926B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision technology, and in particular relates to a zero-shot image classification method, system, device and medium based on collaborative learning. Background Technology
[0002] Currently, zero-shot learning is a branch of machine learning, defined as the absence of training data for certain classes during the model's training phase. These classes are called unseen classes, while the classes that appear during training are called seen classes. Existing image datasets have a fixed number of classes; for example, even the ImageNet dataset, which has the most classes, only has about 30,000 common classes. However, the number of classes in the real world is far greater than this. Therefore, applying zero-shot learning to image classification tasks is essential. Zero-shot learning can not only reduce the amount of training data but also help the model better scale to test data containing entirely new classes.
[0003] Zero-shot image classification methods can be broadly categorized into two types. One type learns a visual-to-semantic mapping of visible classes from the training data, then applies this learned mapping to the classification of unseen classes during the testing phase, emphasizing the learning of the mapping function from image features to semantic features. The other type learns the visual-to-semantic mapping of visible classes, then uses the semantic features of unseen classes to generate features for those classes, addressing the zero-shot problem by completing the training set, focusing on generating features for unseen classes. Both methods have their advantages, but both share a drawback: during the training phase, they utilize only one feature—either semantic or image features—completely neglecting the complementary information between the two features during classification, thus failing to leverage this complementary information for better image classification.
[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows:
[0005] Existing zero-shot image classification methods lack utilization of the complementary information of the two features during the classification phase. Summary of the Invention
[0006] To address the problems existing in the prior art, this invention provides a zero-shot image classification method, system, device, and medium based on collaborative learning.
[0007] This invention is implemented as follows: a zero-shot image classification method based on collaborative learning, wherein the zero-shot image classification method based on collaborative learning includes:
[0008] Establish a dataset; train an image encoder and a semantic encoder; extract image and semantic codes; divide the dataset into two subsets; train classifier 1 and classifier 2 using the image and semantic codes of the two subsets respectively; use classifier 1 and classifier 2 to predict the test data, and divide the test data into the two subsets according to the ranking of the scores; determine whether the subsets have been updated; repeat the division of the two subsets and training of the classifiers until the subsets are no longer updated, then stop the loop; use the finally obtained classifier 1 and classifier 2 to make a final prediction on the test data.
[0009] Furthermore, the dataset includes a training dataset and a test dataset;
[0010] The process of dividing the training dataset into two sub-datasets involves randomly and equally dividing the training dataset into two parts.
[0011] The repeated division of the dataset into two sub-datasets means that in each iteration, the training dataset that has been classified and has included test data is randomly divided into two equal parts.
[0012] Furthermore, the specific process of training the image encoder and semantic encoder is as follows:
[0013] An image encoder is established and trained using the aforementioned training dataset;
[0014] A semantic encoder is established, and supervised training is performed using the image encoder.
[0015] The image encoder consists of four neural network layers: a 2048×1024 fully connected layer, a ReLU activation layer, a 1024×2048 fully connected layer, and a ReLU activation layer. The semantic encoder also consists of four neural network layers: a 4096×1024 fully connected layer, a ReLU activation layer, a 1024×2048 fully connected layer, and a ReLU activation layer.
[0016] Furthermore, the extraction of image encoding and semantic encoding involves inputting the 2048-dimensional image features extracted from the training data based on the pre-trained ResNet50 network into the image encoder for encoding, resulting in 1024-dimensional image encoding; and concatenating the 2048-dimensional semantic features extracted from the training data based on the pre-trained ResNet50 network with 1024-dimensional random noise, and then inputting this into the semantic encoder for encoding, resulting in 1024-dimensional semantic encoding.
[0017] Furthermore, the image encoder is trained using a reconstruction loss function, calculated as follows:
[0018]
[0019] In the formula, p represents the image features of the training data. These are the image features reconstructed after the image encoder.
[0020] The semantic encoder is trained using a reconstruction loss function, and its semantic encoding is supervised using the image encoding. The calculation formula is as follows:
[0021]
[0022] In the formula, 'a' represents the semantic features of the training data. represents the semantic features reconstructed after the semantic encoder; e represents the semantic encoding, and f represents the image encoding.
[0023] Furthermore, the training classifier one and classifier two are trained by using the image encoding of the first subset of the two subsets of the dataset to train classifier one, and by using the semantic encoding of the second subset of the dataset to train classifier two.
[0024] The trained classifiers 1 and 2 are used to classify the test data in the test dataset to obtain the classification confidence scores of the test data. The classification confidence scores are sorted from largest to smallest, and the confidence score of each data point is taken as the larger value of the scores given by classifiers 1 and 2. The top 20% of the data with the highest confidence scores are taken as reliable data. The test data with the classification confidence score of classifier 1 are assigned to subset 2, and the test data with the classification confidence score of classifier 2 are assigned to subset 1, thus completing the re-division of the two subsets.
[0025] There are two classifiers, each scoring the same data point; these are the outputs of the fully connected layer. The maximum score for each class is taken as the confidence score for that data point. This score could be from either classifier 1 or classifier 2. The data is then sorted according to their confidence scores, and the top 20% are assigned to subsets 1 and 2. Data points scored by classifier 1 are assigned to dataset 2, and data points scored by classifier 2 are assigned to dataset 1.
[0026] Furthermore, the final prediction is the maximum confidence score obtained by averaging the classification scores of the final classifiers one and two on the test data, which is taken as the final classification result of the test data.
[0027] Another object of the present invention is to provide a zero-shot image classification system based on collaborative learning that implements the aforementioned zero-shot image classification method based on collaborative learning, the zero-shot image classification system based on collaborative learning comprising:
[0028] Build modules are used to construct datasets and train image encoders and semantic encoders;
[0029] The encoding module is used to extract image encoding and semantic encoding using the image encoder and semantic encoder;
[0030] The training module is used to train classifier one and classifier two using image encoding and semantic encoding of two subsets of the dataset;
[0031] The prediction module is used to predict the test data using the classifier one and the classifier two.
[0032] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the zero-shot image classification method based on collaborative learning.
[0033] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the zero-shot image classification method based on collaborative learning.
[0034] Another objective of this invention is to provide an information data processing terminal for implementing the zero-shot image classification system based on collaborative learning.
[0035] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:
[0036] First, addressing the technical problems existing in the prior art and the difficulty in solving them, this paper closely analyzes, in conjunction with the technical solution to be protected by this invention and the results and data obtained during the research and development process, how the technical solution of this invention solves the technical problems, and the inventive technical effects brought about by solving these problems. The specific description is as follows:
[0037] This invention trains two classifiers using image features and semantic features respectively through collaborative training. The prediction results of one classifier on test data are added to the training data of the other classifier. This method of mixing the data from the two classifiers fuses complementary knowledge from image features and semantic features, ultimately resulting in two classifiers that can recognize both image and semantic features. These two classifiers complement each other, improving the accuracy of ordinary zero-shot classification methods in classifying categories.
[0038] Second, considering the technical solution as a whole or from a product perspective, the technical effects and advantages of the technical solution to be protected by this invention are specifically described as follows:
[0039] This invention employs two classifiers that learn from each other's knowledge to achieve a complementary effect. By fully utilizing two features, it helps ordinary zero-shot image classification methods better identify unseen classes, thereby improving the classification accuracy of the classification network.
[0040] Third, as supplementary evidence of the inventive step of the claims of this invention, it is also reflected in the following important aspects:
[0041] Does the technical solution of this invention solve a technical problem that people have long desired to solve but have never been able to successfully address?
[0042] This invention ingeniously combines image features and semantic features to train two classifiers, fusing the feature information extracted from both. This solves the technical problem of existing methods using only one feature, improves the classification accuracy of zero-sample images, and better identifies visible and unseen classes. Attached Figure Description
[0043] Figure 1 This is a flowchart of a zero-shot image classification method based on collaborative learning provided in an embodiment of the present invention;
[0044] Figure 2 This is a schematic diagram of the image encoder and semantic encoder provided in the embodiments of the present invention;
[0045] Figure 3 These are images of the dataset provided in the embodiments of the present invention: (a) Image 1, (b) Image 2, (c) Image 3, and (d) Image 4.
[0046] Figure 4 This is a schematic diagram of the image encoder provided in an embodiment of the present invention;
[0047] Figure 5 This is a schematic diagram of the semantic encoder provided in an embodiment of the present invention. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0049] To enable those skilled in the art to fully understand how the present invention is specifically implemented, this section provides an explanatory description of the embodiments that expand upon the technical solutions of the claims.
[0050] like Figure 1 As shown, the zero-shot image classification method based on collaborative learning provided in this embodiment of the invention includes:
[0051] Establish a dataset; train an image encoder and a semantic encoder; extract image and semantic codes; divide the dataset into two subsets; train classifier 1 and classifier 2 using the image and semantic codes of the two subsets respectively; use classifier 1 and classifier 2 to predict the test data, and divide the test data into the two subsets according to the ranking of the scores; determine whether the subsets have been updated; repeat the division of the two subsets and training of the classifiers until the subsets are no longer updated; use classifier 1 and classifier 2 obtained from the final iteration to make a final prediction on the test data.
[0052] In this embodiment of the invention, the score for each test sample is the maximum score of each class output by the classification network.
[0053] In an embodiment of the present invention, 经过前两层的 The fully connected layer and ReLU activation layer are encoded, without passing through all 4 layers of the network.
[0054] In this embodiment of the invention, the reconstructed image features are used to calculate Loss1 for training the image encoder. During the final encoding extraction, only the first two layers of the encoder are used. The semantic encoder is processed similarly. In this embodiment of the invention, the reconstructed image features are used to calculate Loss2 for training the semantic encoder.
[0055] In this embodiment of the invention, two classifiers score the same data, which is the output of the fully connected layer. The maximum score for each class is taken as the confidence score of this data; this score could be the score given by classifier 1 or classifier 2. Then, the data are sorted according to their confidence scores, and the top 20% are assigned to subsets 1 and 2. Data with confidence scores from classifier 1 are assigned to dataset 2, and data with confidence scores from classifier 2 are assigned to dataset 1.
[0056] Furthermore, the specific process of the zero-shot image classification method based on collaborative learning is as follows:
[0057] Step 1: Create a zero-shot image classification dataset;
[0058] Step 2: Build an image encoder and train it using the image data from Step 1;
[0059] Step 3: Build a semantic encoder and use the image encoder from Step 2 for supervised training;
[0060] Step 4: Input the images from the zero-shot image classification dataset in Step 1 into the image encoder in Step 2 for encoding, and input the semantic information from the zero-shot image classification dataset in Step 1 into the semantic encoder in Step 3 for encoding. Each image will yield an image code and a semantic code;
[0061] Step 5: Divide the training data of the zero-shot image classification dataset in Step 1 into Sub-dataset 1 and Sub-dataset 2;
[0062] Step 6: Train a classifier 1 using the image encodings from subset 1 in Step 5;
[0063] Step 7: Train a classifier 2 using the semantic encoding of subset 2 from Step 5;
[0064] Step 8: Use classifier 1 and classifier 2, which were trained in steps 6 and 7, to classify the test data in the test dataset;
[0065] Step 9: Select the test data finally determined by classifiers 1 and 2 in Step 8. Add the test data determined by classifier 1 to subset 2, and add the test data determined by classifier 2 to subset 1.
[0066] Step 10: Repeat steps 6 through 9 until the amount of data in datasets 1 and 2 remains unchanged.
[0067] Step 11: Test the test data using the final classifiers 1 and 2, and evaluate the results.
[0068] like Figure 2 As shown, step two involves converting the image data from step one into image features through a pre-trained network before inputting it into the image encoder for training.
[0069] Furthermore, in step two, the image encoder consists of four neural network layers: a fully connected layer, an activation layer, another fully connected layer, and another activation layer. The image encoder is trained using a reconstruction loss function. The calculation formula is as follows:
[0070]
[0071] Where p represents the image features of images in dataset 1. These are the image features reconstructed after the image encoder.
[0072] Furthermore, in step three, the semantic encoder consists of four neural network layers: a fully connected layer, an activation layer, another fully connected layer, and another activation layer. In addition to using a reconstruction loss function, the semantic encoder's training also utilizes image encoding obtained from an image encoder to supervise the semantic encoding process. The calculation formula is as follows:
[0073]
[0074] Where 'a' represents the semantic features of the images in dataset 2. These are the semantic features reconstructed after the semantic encoder. e is the semantic code of the image, and f is the image code of the image.
[0075] Furthermore, in step four, the input to the semantic encoder is a semantic feature concatenated with random noise, and the output is its corresponding semantic code.
[0076] Furthermore, in step five, the training dataset is randomly divided into two equal parts. During each iteration, the already classified data is again randomly divided into two equal parts.
[0077] Furthermore, in steps six and seven, a classifier is trained using a standard zero-shot classification method.
[0078] Furthermore, in step nine, the classification confidence scores of the test samples by classifiers 1 and 2 are sorted from largest to smallest, and the best confidence score is selected. The test data with the classification confidence score of classifier 1 is assigned to dataset 2, and the test data with the classification confidence score of classifier 2 is assigned to dataset 1.
[0079] Furthermore, in step ten, the loop stops when there is no data available in the test dataset or no test data that meets the required confidence threshold.
[0080] Furthermore, in step eleven, the maximum confidence score obtained by averaging the classification scores of the test image by the final classifiers 1 and 2 is taken as the final classification result of the test image.
[0081] To demonstrate the inventiveness and technical value of the technical solution of this invention, this section provides specific product or related technology application examples of the technical solution claimed.
[0082] The main steps of the technical solution of this invention are as follows:
[0083] Step 1: Use the CUB-200 bird dataset as the zero-shot dataset, which contains 11,788 images across 200 classes. Select 7,057 images (150 classes) as the training dataset, and the remaining 4,731 images (50 classes) as the test set. (Example follows.) Figure 3 As shown, 2048-dimensional image features of all images in the dataset are extracted using a pre-trained ResNet50 network on ImageNet, with each image corresponding to a 1024-dimensional semantic feature.
[0084] Step 2: Build an image encoder consisting of a 4-layer neural network, with a structure of 2048×1024 fully connected layers and activation layers, and 1024×2048 fully connected layers and activation layers, as shown below. Figure 4As shown, the image encoder takes 2048-dimensional image features p as input and outputs 2048-dimensional reconstructed features. The intermediate encoding is a 1024-dimensional image encoding f. The image encoder is trained using a reconstruction loss function. The formula is as follows:
[0085]
[0086] Step 3: Build a semantic encoder consisting of a 4-layer neural network, with a 4096×1024 fully connected layer and activation layer, and a 1024×2048 fully connected layer and activation layer, as shown below. Figure 5 As shown, the image encoder takes as input a concatenation of 2048-dimensional semantic features 'a' and 1024-dimensional random noise, and outputs 2048-dimensional reconstructed features. The intermediate encoding is a 1024-dimensional semantic code e. In addition to using a reconstruction loss function, the semantic encoder's training also uses image encodings obtained from the image encoder to supervise the semantic encoding process. The calculation formula is as follows:
[0087]
[0088] Step 4: Encode the image features of the training dataset from Step 1 using the image encoder obtained in Step 2. Encode the semantic features of the training dataset from Step 1 using the semantic encoder obtained in Step 3. Each image will receive both a 1024-dimensional image encoding and a 1024-dimensional semantic encoding.
[0089] Step 5: Randomly and equally divide the training data of the zero-shot image classification dataset in Step 1 into Subset 1 and Subset 2.
[0090] Step 6: Train a CLS-WGAN zero-shot image classifier using the image encodings from sub-dataset 1 in Step 5.
[0091] Step 7: Train a CLS-WGAN zero-shot image classifier 2 using the semantic encoding of subset 2 from Step 5.
[0092] Step 8: Use classifier 1 and classifier 2, which were trained in steps 6 and 7, to score the data in the test dataset and obtain the classification confidence of each test image through the classifiers.
[0093] Step Nine: Sort the data scored by classifiers 1 and 2 in Step Eight. For each data point, take the larger confidence score from both classifiers 1 and 2, and select the top 20% of data with the highest confidence scores as reliable data. Add the scores of classifier 1 as test data for its confidence score to subset 2, and add the scores of classifier 2 as test data for its confidence score to subset 1. Remove these data from the test data.
[0094] Step 10: Repeat steps 6 to 9 until the amount of data in datasets 1 and 2 no longer changes or until there are no more images in the test dataset.
[0095] Step 11: Take the zero-sample image classifiers 1 and 2 obtained from the final loop and make the final prediction on the test dataset. Take the average score of the two classifiers on the test images as the classification result of the test data.
[0096] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.
[0097] The embodiments of the present invention have achieved some positive results during the research and development or use process, and have indeed great advantages compared with the prior art. The following content describes them in conjunction with the data, charts and other information of the experimental process.
[0098] After testing, the present invention achieved a Top-1 accuracy of 60.2% for the visible class, a Top-1 accuracy of 50.1% for the unseen class, and a harmonic accuracy of 54.7% on the CUB-200 bird dataset. These results are higher than the 57.7%, 43.7%, and 49.7% obtained by the underlying CLS-WGAN method, fully demonstrating the improvement of the present invention over ordinary zero-shot image classification methods, and its ability to better identify the visible and unseen classes.
[0099] It should be noted that this invention builds upon conventional zero-shot image classification methods by training two zero-shot image classifiers through collaborative learning. The two classifiers expand their training data by exchanging data they deem reliable, thereby fusing classification information from both types of features. Ultimately, this results in two classifiers that simultaneously recognize image features and semantic features. This helps conventional zero-shot image classification methods to perform better zero-shot image classification.
[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A zero-shot image classification method based on collaborative learning, characterized in that, The zero-shot image classification method based on collaborative learning includes: Establish a dataset; train an image encoder and a semantic encoder; extract image and semantic codes; divide the dataset into two subsets; train classifier 1 and classifier 2 using the image and semantic codes of the two subsets respectively; use classifier 1 and classifier 2 to predict test data, sort the data according to the maximum classification score obtained for each test data, select the top-ranked data, assign pseudo-labels to them, and allocate them to subsets; determine if the subsets have been updated; repeat the process of dividing the dataset into two subsets and training the classifiers until the subsets are no longer updated, then stop the loop; use the final classifier 1 and classifier 2 to make a final prediction on the test data. The training classifier one and classifier two are trained by using the image encoding of the first subset of the two subsets of the dataset, and by using the semantic encoding of the second subset of the dataset. The test data in the test dataset is classified using the trained classifier 1 and classifier 2 to obtain the classification confidence of the test data. The classification confidence is sorted from largest to smallest. The confidence of each data is taken as the larger value of the scores given by classifier 1 and classifier 2. The top 20% of the data with high confidence are taken as reliable data. The test data with the top 20% classification confidence of classifier 1 are assigned to subset 2, and the test data assigned to classifier 2 are assigned to subset 1, thus completing the re-division of the two subsets. Data allocation involves adding the prediction results of one classifier on the test data to the training data of another classifier.
2. The zero-shot image classification method based on collaborative learning as described in claim 1, characterized in that, The dataset includes a training dataset and a test dataset; The process of dividing the training dataset into two sub-datasets involves randomly and equally dividing the training dataset into two parts. The repeated division of the dataset into two sub-datasets means that in each iteration, the training dataset that has been classified and has included test data is randomly divided into two equal parts.
3. The zero-shot image classification method based on collaborative learning as described in claim 1, characterized in that, The specific process of training the image encoder and semantic encoder is as follows: An image encoder is established and trained using the aforementioned training dataset; A semantic encoder is established, and supervised training is performed using the image encoder. The image encoder consists of four neural network layers: a 2048×1024 fully connected layer, a ReLU activation layer, a 1024×2048 fully connected layer, and a ReLU activation layer. The semantic encoder also consists of four neural network layers: a 4096×1024 fully connected layer, a ReLU activation layer, a 1024×2048 fully connected layer, and a ReLU activation layer.
4. The zero-shot image classification method based on collaborative learning as described in claim 1, characterized in that, The extracted image encoding and semantic encoding involve inputting the 2048-dimensional image features extracted from the training data based on the pre-trained ResNet50 network into the first two fully connected layers and ReLU activation layers of the image encoder for encoding, resulting in 1024-dimensional image encoding; and concatenating the 2048-dimensional semantic features extracted from the training data based on the pre-trained ResNet50 network with 1024-dimensional random noise, then inputting this into the semantic encoder for encoding, resulting in 1024-dimensional semantic encoding.
5. The zero-shot image classification method based on collaborative learning as described in claim 1, characterized in that, The image encoder is trained using a reconstruction loss function, calculated as follows: In the formula, Image features for training data, These are the 2048-dimensional image features reconstructed after the image encoder. The semantic encoder is trained using a reconstruction loss function and supervised by the image encoding. The calculation formula is as follows: In the formula, For the semantic features of the training data, The 2048-dimensional semantic features reconstructed after the semantic encoder; For semantic encoding, Encode the image.
6. The zero-shot image classification method based on collaborative learning as described in claim 1, characterized in that, The final prediction is the maximum confidence score obtained by averaging the classification scores of the final classifiers one and two on the test data, which is taken as the final classification result of the test data.
7. A zero-shot image classification system based on collaborative learning, implementing the zero-shot image classification method based on collaborative learning as described in any one of claims 1-6, characterized in that, The zero-shot image classification system based on collaborative learning includes: Build modules are used to construct datasets and train image encoders and semantic encoders; The encoding module is used to extract image encoding and semantic encoding using the image encoder and semantic encoder; The training module is used to train classifier one and classifier two using image encoding and semantic encoding of two subsets of the dataset; The prediction module is used to predict the test data using the classifier one and the classifier two.
8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the zero-shot image classification method based on collaborative learning as described in any one of claims 1-6.
9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the zero-shot image classification method based on collaborative learning as described in any one of claims 1-6.