An image retrieval method, system, device and medium for enhancing fine-grained object retrieval performance

Abstract

The present disclosure relates to the technical field of image retrieval, and provides an image retrieval method, system, device and medium for enhancing fine-grained object retrieval performance. The method processes original images through a multi-scale image interception technology to generate target images and a slice set. These images are input into an encoder to construct an image vector library. Meanwhile, a multi-modal large model performs semantic analysis on the images and slices to generate text descriptions and construct a text vector library. In the retrieval stage, visual similarity matching is performed based on the image vectors, and text similarity matching is performed based on the text vector library. The results are integrated to obtain a candidate set. Finally, through weighted fusion and sorting, the final retrieval result is obtained. The present disclosure significantly improves the accuracy of fine-grained object retrieval. By combining visual and text information, not only is the comprehensiveness of the retrieval enhanced, but also the relevance of the results is improved, enabling users to more quickly and accurately find the required information.

Description

Technical Field

[0001] This disclosure belongs to the field of image retrieval technology, and in particular relates to an image retrieval method, system, device and medium for enhancing fine-grained object retrieval performance. Background Technology

[0002] With the continuous advancement of computer technology, the management and application of image data in search systems are becoming increasingly important. Introducing image information into question-answering systems and computer vision tasks can significantly improve the accuracy and intelligence of these systems; however, some key challenges still exist in current image information retrieval practices.

[0003] Currently, commonly used image retrieval methods mainly rely on the embedding vectors generated by image coding models to achieve matching through "image-to-image search". This method is difficult to support the function of "text-to-image search", that is, finding relevant images based on text descriptions. In addition, due to the different sizes of images and the fact that the target object may only occupy a small part of the image, the information proportion of the object in the whole image is small, which reduces the similarity matching effect between the overall images and easily leads to the omission of relevant images.

[0004] Therefore, proposing an image retrieval method to enhance the performance of fine-grained object retrieval, aiming to improve the accuracy and semantic understanding of image retrieval, is a feasible and preferred approach. Summary of the Invention

[0005] To address the aforementioned issues, this disclosure provides an image retrieval method, system, device, and medium that enhances the performance of fine-grained object retrieval. By introducing an object-based multi-scale slicing mechanism and combining it with the Paligemma model to generate image semantic descriptions, a multi-level understanding of image content is achieved.

[0006] In a first aspect, this disclosure provides an image retrieval method to enhance the performance of fine-grained object retrieval, the method comprising,

[0007] The original image is processed using multi-scale image cropping technology to obtain the target image and a set of target image slices;

[0008] The target image and the set of target image slices are input into the image encoder to establish an image vector library;

[0009] A multimodal large model is used to perform semantic analysis on the target image and the set of target image slices to generate corresponding text descriptions. Each text description is then vector-encoded to establish a text vector library.

[0010] The query vector is visually similar to the image vector library and semantically similar to the text vector library to obtain a set of visual candidate images and a set of semantic candidate images, and a visual similarity score and a semantic similarity score are generated.

[0011] The visual similarity score and the semantic similarity score are weighted and fused to obtain the final similarity score. The visual candidate image set and the semantic candidate image set are then sorted according to the final similarity score to obtain the final retrieval result.

[0012] Furthermore,

[0013] The original image is processed using multi-scale image cropping techniques to obtain the target image and a set of target image slices, specifically including:

[0014] The original image is resized uniformly to obtain the target image;

[0015] The target image is identified using an image segmentation model to obtain each target object and its corresponding target location information;

[0016] The target image is cropped at multiple scales according to preset size parameters to obtain an initial set of image slices.

[0017] Based on the target object and its target location information, the object inclusion relationship between each initial image slice is determined, and the initial image slice set is deduplicated according to the object inclusion relationship to obtain the target image slice set.

[0018] Furthermore,

[0019] The target image and the set of target image slices are input into an image encoder to establish an image vector library, specifically including:

[0020] The target image and the set of target image slices are input into the image encoder to extract the corresponding visual features;

[0021] The extracted visual features are fused to establish the image vector library.

[0022] Furthermore,

[0023] The query vector is visually similar to the image vector library to obtain a set of visual candidate images, and a visual similarity score is generated, specifically including:

[0024] The query vector is used to search the image vector database to obtain an initial set of candidate images;

[0025] Calculate the visual similarity score between each candidate image in the initial candidate image set and the query vector, and determine the visual candidate image set based on each visual similarity score.

[0026] Furthermore,

[0027] The query vector is semantically matched with the text vector library to obtain a set of semantic candidate images, and a semantic similarity score is generated, specifically including:

[0028] The query vector is used to search the text vector database to obtain an initial set of candidate text descriptions.

[0029] The actual similarity between the query vector and each text description in the text vector library is determined using a text similarity algorithm, and two initial candidate text description sets are determined based on the actual similarity.

[0030] The initial candidate text description set of the first type and the initial candidate text description set of the second type are deduplicated to obtain the actual recall result;

[0031] The semantic similarity score is calculated on the actual recall results using a text rerank model, and the semantic candidate image set is determined based on the semantic similarity score.

[0032] Furthermore,

[0033] The visual similarity score and the semantic similarity score are weighted and fused to obtain a final similarity score. Based on this final similarity score, the visual candidate image set and the semantic candidate image set are sorted to obtain the final retrieval result, specifically including:

[0034] Obtain the inverse ranking of each final candidate image in the visual candidate image set and the inverse ranking of the text in the semantic candidate image set;

[0035] The final similarity score is determined based on the inverse ranking of the images and the inverse ranking of the text.

[0036] Secondly, based on the same inventive concept, this disclosure provides an image retrieval system that enhances fine-grained object retrieval performance.

[0037] The system includes:

[0038] The multi-scale image cropping module processes the original image based on multi-scale image cropping technology to obtain the target image and a set of target image slices;

[0039] The image feature extraction module inputs the target image and the set of target image slices into the image encoder to establish an image vector library;

[0040] The image semantic description generation module uses a multimodal large model to perform semantic analysis on the target image and the target image slice set, generates corresponding text descriptions, and encodes each text description into a vector to establish a text vector library.

[0041] The first retrieval module performs visual similarity matching between the query vector and the image vector library, and semantic similarity matching between the query vector and the text vector library, to obtain a set of visual candidate images and a set of semantic candidate images, and generates visual similarity scores and semantic similarity scores.

[0042] The secondary retrieval module performs a weighted fusion of the visual similarity score and the semantic similarity score to obtain a final similarity score, and sorts the visual candidate image set and the semantic candidate image set according to the final similarity score to obtain the final retrieval result.

[0043] Thirdly, this disclosure also provides an electronic device, including at least one processor and at least one memory electrically connected;

[0044] The memory is electrically connected to the processor, wherein the memory stores instructions executable by at least one of the processors, the instructions being executed by at least one of the processors to enable at least one of the processors to perform any of the image retrieval methods described above for enhancing fine-grained object retrieval performance.

[0045] Fourthly, this disclosure also provides a computer storage medium in which a computer program is stored;

[0046] When the computer program is executed by the processor, it implements any of the image retrieval methods described above that enhance the performance of fine-grained object retrieval.

[0047] Fifthly, this disclosure also provides a computer program product, which is stored in at least one storage medium;

[0048] The computer program product includes several instructions for causing at least one electronic device to execute any of the image retrieval methods described above that enhance fine-grained object retrieval performance.

[0049] Compared with the prior art, this disclosure has the following advantages:

[0050] 1. By using an image segmentation model combined with minimum size constraints, we can accurately identify and extract each object in the image, ensuring the integrity of each object. At the same time, we adopt a multi-scale cutting strategy to obtain image content from different granularities, thereby enhancing the ability to capture local details.

[0051] 2. Input the original image and its multiple slices into the image encoder to extract visual features at different levels and fuse them to form a more representative comprehensive image vector representation, thereby improving the accuracy of image similarity calculation.

[0052] 3. By utilizing multimodal large models such as Paliemema, high-quality text descriptions are automatically generated for each image, establishing a mapping relationship between the image and its semantic information, and thus constructing an image semantic information database, which greatly expands the application scenarios of image retrieval.

[0053] Other features and advantages of this disclosure will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating an embodiment of the image retrieval method for enhancing fine-grained object retrieval performance is shown. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0057] Figure 1 An image retrieval method for enhancing fine-grained object retrieval performance according to an embodiment of this disclosure is illustrated. For example... Figure 1 As shown, the image retrieval method for enhancing fine-grained object retrieval performance according to embodiments of this disclosure includes the following steps:

[0058] S1, The original image is processed based on multi-scale image cropping technology to obtain the target image and a set of target image slices;

[0059] In the embodiments of this disclosure, step S1 specifically includes:

[0060] S11, The original image is uniformly resized to obtain the target image;

[0061] S12, the target image is identified using an image segmentation model to obtain each target object and the target location information corresponding to each target object;

[0062] S13, perform multi-scale cropping on the target image according to preset size parameters to obtain an initial image slice set;

[0063] S14, determine the object inclusion relationship between each initial image slice based on each target object and its target location information, and perform deduplication processing on the initial image slice set according to the object inclusion relationship to obtain the target image slice set.

[0064] In practice, the specific process of cropping based on multi-scale image cropping technology includes:

[0065] 1. Standardize the original input image size to 1024x1024 to obtain the target image;

[0066] 2. Based on the image segmentation model, obtain the specific location information of each target object in the target image;

[0067] 3. Using image sizes of 512x512, 224x224, 112x112, and 32x32 with a step size of 16, crop the image to obtain target image slices of each size.

[0068] 4. Perform deduplication on each target image slice, retaining target image slices that are semantically independent and have irreplaceable content, and removing target image slices that are completely contained by other target image slices or can be covered by combinations of other target image slices.

[0069] For a target image slice cropped to any image size:

[0070] 1) Count the number of objects in each slice of the target image;

[0071] 2) Traverse all the segmented target image slices;

[0072] 3) Using the segmentation model results for each target image slice, count the number of target objects contained within it, as well as the category and location information of each target object.

[0073] 4) Determine the object inclusion relationship between each target image slice. If all target objects in target image slice A can be found in target image slice B with the same category and completely contained in the same location, then target image slice B is considered to completely contain the content of target image slice A. At this time, it can be determined that target image slice A is a "subset" of target image slice B, and target image slice A can be deleted.

[0074] 5) Pairwise comparison for deduplication: Perform pairwise comparisons on all target image slices: If target image slice B contains all target objects of target image slice A → delete target image slice A; if target image slice A contains all target objects of target image slice B → delete target image slice B; if they do not contain each other → keep both target image slices.

[0075] 6) Interval merging optimization: There is a set of continuously segmented target image slices: image i to image j; if the combined content of image i and image j can cover all target objects in the intermediate target image slices, that is, the combined content of image i and image j contains all target objects appearing from image i to image j; then image i+1 to image j-1 can be deleted, keeping only image i and image j.

[0076] 7) Finally, retain the target image slice set. After the above steps, a set of target image slices with no redundancy and complementary information is obtained. These images will be used for subsequent feature extraction and index construction.

[0077] For example: Suppose we have the following three target image slices: Target image slice A: contains the object {cat}; Target image slice B: contains the objects {cat, window}; Target image slice C: contains the objects {window, table}; According to the rules: Target image slice A is completely contained by target image slice B → Delete target image slice A; Target image slice B and target image slice C have no mutual containment relationship → Both are retained; The final retained target image slices are: B and C.

[0078] S2, Input the target image and the set of target image slices into the image encoder to establish an image vector library;

[0079] In the embodiments of this disclosure, step S2 specifically includes:

[0080] S21, input the target image and the set of target image slices into the image encoder to extract the corresponding visual features;

[0081] S22, the extracted visual features are fused to establish the image vector library.

[0082] In this embodiment, the original image and the obtained set of target image slices without redundancy and with complementary information are fed into the image encoder of the Paligemma model to form a comprehensive image vector representation containing multi-level information. In specific implementation, this is the mapping relationship from the target image slices to the original image.

[0083] S3, use a multimodal large model to perform semantic analysis on the target image and the target image slice set, generate corresponding text descriptions, encode each text description into a vector, and establish a text vector library;

[0084] In this embodiment, in order to extract semantic information from images more efficiently, an advanced Paligemma model is used to process the original image and its multi-scale slices (a set of small images without redundancy and with complementary information), automatically generating detailed image descriptions. This allows for understanding the image content from both visual and semantic dimensions, providing rich text information support for the subsequent retrieval process. Finally, the semantic information is vector-encoded, and a mapping relationship between semantics and images is established. The Paligemma model is uniformly selected for image encoding and text description generation, which ensures alignment between image and text semantics.

[0085] S2 and S3 have generated the original image, as well as image vector libraries and text vector libraries for multi-size slices. By searching for the image or semantic information of the multi-size slices, the original image information can be mapped.

[0086] S4, perform visual similarity matching between the query vector and the image vector library, and semantic similarity matching between the query vector and the text vector library to obtain a set of visual candidate images and a set of semantic candidate images, and generate visual similarity scores and semantic similarity scores.

[0087] In the embodiments of this disclosure, step S4 specifically includes:

[0088] S41, using the query vector, a search is performed in the image vector library to obtain an initial candidate image set;

[0089] S42, calculate the visual similarity score between each candidate image in the initial candidate image set and the query vector, and determine the visual candidate image set based on each visual similarity score;

[0090] S43, using the query vector, a search is performed in the text vector library to obtain an initial set of candidate text descriptions;

[0091] S44, use a text similarity algorithm to determine the actual similarity between the query vector and each text description information in the text vector library, and determine two initial candidate text description sets based on each actual similarity;

[0092] S45, perform deduplication processing on the first type of initial candidate text description set and the second type of initial candidate text description set to obtain the actual recall result;

[0093] S46, use the text rerank model to calculate the semantic similarity score of the actual recall result, and determine the semantic candidate image set based on the semantic similarity score.

[0094] S5, perform weighted fusion of the visual similarity score and the semantic similarity score to obtain the final similarity score, and sort the visual candidate image set and the semantic candidate image set according to the final similarity score to obtain the final retrieval result;

[0095] In the embodiments of this disclosure, step S5 specifically includes:

[0096] S51, obtain the inverse ranking of each final candidate image in the visual candidate image set and the inverse ranking of the text in the semantic candidate image set;

[0097] S52, determine the final similarity score based on the inverse ranking of the image and the inverse ranking of the text.

[0098] In practice, the query vector can be text data or image data;

[0099] 1) Text-based image search: When the query vector is text, the query text is converted into a vector representation using a text encoding model;

[0100] 2) Image search: When the query vector is an image, the query image is converted into a vector representation through an image encoding model, and text description information is generated using the Paligemma model to prepare for subsequent text retrieval.

[0101] 3) Image similarity matching: Using the above text vectors (text search image as text vector) / image vectors (image search image as image vector), the most relevant image information is searched in the pre-built image vector library; for each retrieved original image (which may contain multiple slices of different sizes), the visual similarity score between the original image and the query vector is calculated according to the similarity calculation formula.

[0102] The similarity calculation formula is as follows:

[0103]

[0104] in, It is the average value of the slice similarity; The calculation uses variance (or dispersion) to measure the difference in similarity between slices. If the similarity of all slices is very close, the variance is small; conversely, if the similarity difference is large, the variance is large.

[0105] In the specific implementation process, when calculating the comprehensive similarity of the original images based on multi-size slices, the following factors need to be considered:

[0106] The impact of individual slice similarity: Slices of each size contribute to the final similarity.

[0107] The impact of the number of slices: Finding slices of more sizes should increase the overall similarity.

[0108] Balancing mechanism: Ensures that in certain situations (e.g., when all slices have low similarity), the overall similarity is not incorrectly increased due to a large number of slices.

[0109] 4) Text-based search: Semantic search is performed in the text vector library to find relevant text information that is most similar to the input text semantics (for text-to-image search, the input text semantics are the original query text; for image-to-image search, the input text semantics are the text description information generated by the Paligemma model). Simultaneously, text similarity evaluation is performed using a text similarity algorithm (such as the text similarity mechanism in Elasticsearch) to assess the similarity between the query text and the text description information. The text description information obtained from both search methods is deduplicated and used as the recall result. Re-ranking: Finally, a text reranking model is applied to re-rank the above recall results to obtain the similarity score of the recall results; finally, according to the similarity calculation formula, the similarity score corresponding to the original image is calculated from the text description dimension.

[0110] 5) A fusion and ranking algorithm for visual similarity scores and semantic similarity scores. For each image, its inverse ranking in the image list and text description list is summed to obtain the fusion score for that image. That is, the fused similarity score of the i-th retrieved image is:

[0111]

[0112] Where k is the weight of the image list sorting, and (1-k) is the weight of the text description list. If searching for images by text, since text information is relatively more accurate, k=0.4, 1-k=0.6, and the sorting of the text information part has a greater weight. If searching for images by image, then k=0.6, and the image part has a greater weight. To sort the i-th image in the image list, Sort the i-th image in the list of text descriptions.

[0113] In this embodiment, by employing the aforementioned object-based multi-scale image cropping and encoding, along with a dual retrieval mechanism combining image and semantic features, the present invention significantly improves the performance of the image retrieval system, particularly excelling in fine-grained object retrieval. Furthermore, this method demonstrates great potential in intelligent question-answering scenarios, enabling more efficient utilization of image information and thereby enhancing the overall effectiveness of the question-answering system.

[0114] Based on the same inventive concept, this disclosure also provides an image retrieval system with enhanced fine-grained object retrieval performance corresponding to the above-described method, including:

[0115] The multi-scale image cropping module processes the original image based on multi-scale image cropping technology to obtain the target image and a set of target image slices;

[0116] The image feature extraction module inputs the target image and the set of target image slices into the image encoder to establish an image vector library;

[0117] The image semantic description generation module uses a multimodal large model to perform semantic analysis on the target image and the target image slice set, generates corresponding text descriptions, and encodes each text description into a vector to establish a text vector library.

[0118] The first retrieval module performs visual similarity matching between the query vector and the image vector library, and semantic similarity matching between the query vector and the text vector library, to obtain a set of visual candidate images and a set of semantic candidate images, and generates visual similarity scores and semantic similarity scores.

[0119] The secondary retrieval module performs a weighted fusion of the visual similarity score and the semantic similarity score to obtain a final similarity score, and sorts the visual candidate image set and the semantic candidate image set according to the final similarity score to obtain the final retrieval result.

[0120] Based on the same inventive concept, this disclosure also provides an electronic device. The electronic device of this disclosure includes at least one processor and at least one memory electrically connected to the processor. The memory stores instructions executable by the at least one processor, which, when executed, enables the at least one processor to perform an image retrieval method as described above that enhances fine-grained object retrieval performance.

[0121] It should be noted that the electrical connections between the various units mentioned above do not necessarily represent the connections between lines. Any indirect connection method can be applied to the embodiments of this disclosure as long as it achieves the purpose of this disclosure.

[0122] Based on the same inventive concept, this disclosure also provides a computer storage medium storing a computer program, which, when executed by a processor, implements an image retrieval method for enhancing fine-grained object retrieval performance as described above.

[0123] Based on the same inventive concept, this disclosure also provides a computer program product stored in at least one storage medium; the computer program product includes several instructions to cause at least one computer device to execute the image retrieval method described above for enhancing fine-grained object retrieval performance.

[0124] Although the present disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure.

Claims

1. An image retrieval method for enhancing fine-grained object retrieval performance, characterized in that, The method includes: The original image is processed using multi-scale image cropping technology to obtain the target image and a set of target image slices; The target image and the set of target image slices are input into the image encoder to establish an image vector library; A multimodal large model is used to perform semantic analysis on the target image and the set of target image slices to generate corresponding text descriptions. Each text description is then vector-encoded to establish a text vector library. The query vector is visually similar to the image vector library and semantically similar to the text vector library to obtain a set of visual candidate images and a set of semantic candidate images, and a visual similarity score and a semantic similarity score are generated. The visual similarity score and the semantic similarity score are weighted and fused to obtain the final similarity score. The visual candidate image set and the semantic candidate image set are then sorted according to the final similarity score to obtain the final retrieval result.

2. The method according to claim 1, characterized in that, The original image is processed using multi-scale image cropping techniques to obtain the target image and a set of target image slices, specifically including: The original image is resized uniformly to obtain the target image; The target image is identified using an image segmentation model to obtain each target object and its corresponding target location information; The target image is cropped at multiple scales according to preset size parameters to obtain an initial set of image slices. Based on the target object and its target location information, the object inclusion relationship between each initial image slice is determined, and the initial image slice set is deduplicated according to the object inclusion relationship to obtain the target image slice set.

3. The method according to claim 2, characterized in that, The target image and the set of target image slices are input into an image encoder to establish an image vector library, specifically including: The target image and the set of target image slices are input into the image encoder to extract the corresponding visual features; The extracted visual features are fused to establish the image vector library.

4. The method according to claim 3, characterized in that, The query vector is visually similar to the image vector library to obtain a set of visual candidate images, and a visual similarity score is generated, specifically including: The query vector is used to search the image vector database to obtain an initial set of candidate images; Calculate the visual similarity score between each candidate image in the initial candidate image set and the query vector, and determine the visual candidate image set based on each visual similarity score.

5. The method according to claim 4, characterized in that, The query vector is semantically matched with the text vector library to obtain a set of semantic candidate images, and a semantic similarity score is generated, specifically including: The query vector is used to search the text vector database to obtain an initial set of candidate text descriptions. The actual similarity between the query vector and each text description in the text vector library is determined using a text similarity algorithm, and two initial candidate text description sets are determined based on the actual similarity. The initial candidate text description set of the first type and the initial candidate text description set of the second type are deduplicated to obtain the actual recall result; The semantic similarity score is calculated on the actual recall results using a text rerank model, and the semantic candidate image set is determined based on the semantic similarity score.

6. The method according to claim 5, characterized in that, The visual similarity score and the semantic similarity score are weighted and fused to obtain a final similarity score. Based on this final similarity score, the visual candidate image set and the semantic candidate image set are sorted to obtain the final retrieval result, specifically including: Obtain the inverse ranking of each final candidate image in the visual candidate image set and the inverse ranking of the text in the semantic candidate image set; The final similarity score is determined based on the inverse ranking of the images and the inverse ranking of the text.

7. An image retrieval system with enhanced fine-grained object retrieval performance, characterized in that, The system includes: The multi-scale image cropping module processes the original image based on multi-scale image cropping technology to obtain the target image and a set of target image slices; The image feature extraction module inputs the target image and the set of target image slices into the image encoder to establish an image vector library; The image semantic description generation module uses a multimodal large model to perform semantic analysis on the target image and the target image slice set, generates corresponding text descriptions, and encodes each text description into a vector to establish a text vector library. The first retrieval module performs visual similarity matching between the query vector and the image vector library, and semantic similarity matching between the query vector and the text vector library, to obtain a set of visual candidate images and a set of semantic candidate images, and generates visual similarity scores and semantic similarity scores. The secondary retrieval module performs a weighted fusion of the visual similarity score and the semantic similarity score to obtain a final similarity score, and sorts the visual candidate image set and the semantic candidate image set according to the final similarity score to obtain the final retrieval result.

8. An electronic device comprising at least one processor and at least one memory electrically connected to each other; The memory is electrically connected to the processor, wherein... The memory stores instructions executable by at least one of the processors, which enable the at least one processor to perform an image retrieval method for enhancing fine-grained object retrieval performance as described in any one of claims 1-6.

9. A computer storage medium, characterized in that, The computer-readable storage medium stores a computer program. When the computer program is executed by the processor, it implements the image retrieval method for enhancing fine-grained object retrieval performance as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product is stored in at least one storage medium; The computer program product includes several instructions for causing at least one electronic device to execute the image retrieval method for enhancing fine-grained object retrieval performance as described in any one of claims 1-6.

Images