Zero sample reference image segmentation method based on hierarchical prompt and directional clue
By introducing hierarchical hints and directional cues in zero-sample reference image segmentation, combined with advanced models such as RAM++, Grounding DINO and HQ-SAM, the poor mask generation and insufficient utilization of spatial position information in the prior art are solved. Problem, high-precision and detailed image segmentation effect are achieved.
Patent Information
- Application Number
- CN202411011071.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-07-26
AI Technical Summary
The existing zero-sample referential image segmentation method performs poorly in precision and handling complex segmentation scenarios, and is prone to generate a large number of redundant and trivial masks. The CLIP model cannot effectively utilize the clear directional clues in the referential expression.
Using a method based on hierarchical prompts and directional cues, the mask of the input image is obtained through the hierarchical prompt mask generation network, combined with visual and text features, and used pre-trained models such as BLIP2 and CLIP for feature extraction and matching, and introduced a spatial rectifier. To improve the location sensitivity of the mask, the Recognize Anything Plus Model (RAM++), Grounding DINO, and Segment Anything in High Quality (HQ-SAM) models are integrated to generate detailed masks.
Significantly improve mask quality, reduce redundancy, enhance fine-grained and accurate image segmentation, especially in occluded and dense scenes, and improve the spatial sensitivity and detailed identification capabilities of the model. .
Smart Images

Figure CN119049057A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image segmentation, and in particular to a zero-sample referential image segmentation method based on hierarchical prompts and directional clues. Background Art
[0002] Referring Image Segmentation aims to accurately segment the target area specified by the natural language description in the image. This process requires precise alignment between the image and the text, as well as a comprehensive understanding of the visual and textual elements. This poses a major challenge in the field of visual language. In addition, generating accurate mask annotations and corresponding natural language descriptions is time-consuming and expensive, and errors in manual annotations may affect the final results. To address these issues, researchers have explored weakly supervised methods. However, this method has poor performance and still relies on high-quality training datasets. In contrast, zero-shot referring image segmentation is a method that can accurately identify the masks most relevant to referring expressions without relying on pixel-level annotations. This process includes mask generation and mask-to-text matching, which is crucial for generating accurate, high-quality masks and exploring visual-textual relationships, making it more attractive and research-worthy.
[0003] Recently, multimodal pre-trained models have shown impressive capabilities in vision and language understanding. Of particular note is the visual language model CLIP, which excels in capturing global similarities between text and images. This is demonstrated by its outstanding performance in various image-level zero-shot tasks, including image retrieval, object detection, and semantic segmentation. However, the CLIP model is trained on image-text pairs, while the task of referential image segmentation requires dense pixel-level predictions. Therefore, it is challenging to directly apply the pre-trained CLIP model to the zero-shot referential image segmentation task that requires fine-grained region-text matching. Moreover, the CLIP model is insensitive to spatial location information, so this information is inevitably reduced during the model's image encoding process. However, referential expressions often contain clear directional cues (e.g., "the woman on the right", "the bread on the table", etc.), which are valuable but easily overlooked.
[0004] The quality of masks generated by the mask generation network plays a crucial role in determining the upper limit performance of this task. In Global-Local CLIP, FreeSOLO is used as the mask generation network, but it performs poorly in terms of accuracy and handling complex segmentation scenarios. TAS uses SAM as the mask generation network to improve the quality of mask generation, but due to insufficient hints, SAM is prone to generate a large number of redundant and trivial masks. Therefore, it is difficult for current zero-shot referential image segmentation schemes to achieve good segmentation performance. Summary of the invention
[0005] The purpose of the present invention is to overcome the problems in the above-mentioned prior art that the mask generation network performs poorly in terms of accuracy and processing complex segmentation scenes and is prone to generate a large number of redundant and trivial masks, and that CLIP cannot effectively utilize the clear directional clues in referential expressions, and to provide a zero-sample referential image segmentation method based on hierarchical cues and directional cues.
[0006] In order to achieve the above object, the present invention provides a zero-sample referential image segmentation method based on hierarchical cues and directional cues, comprising the following steps:
[0007] S1, obtain masks of all object instances in the input image through a hierarchical hint mask generation network;
[0008] S2. Extract comprehensive visual features based on input image I and input text T and focal visual features At the same time, the two are fused to obtain the comprehensive-focal visual feature F V ;
[0009] S3. Get the input text T, generate the title text and negative expression text using the pre-trained model BLIP2, and extract the title features using the CLIP text encoder Negative text features Then, for the input text T, the CLIP text encoder is used to extract comprehensive text features And the focus text features At the same time, the two are combined to obtain the comprehensive-focus text feature F T ;
[0010] S4. Use the pre-trained model CLIP to calculate the cosine similarity between the input image I and the input text T, introduce a spatial rectifier to use the center of the mask as the position information of each mask scheme, and select the mask with the highest matching score in the corresponding direction area.
[0011] Preferably, the hierarchical prompt mask generation network integrates three advanced models: Recognize Anything Plus Model (RAM++), Grounding DINO and Segment Anything in High Quality (HQ-SAM).
[0012] Preferably, the step S1 specifically includes the following steps:
[0013] S101, generating category labels for all object instances using RAM++ according to the input image I as the first layer of prompts;
[0014] S102, using Grounding DINO to generate a target detection frame from the category label as a second-level prompt;
[0015] S103 . HQ-SAM generates segmentation masks for all object instances in the image using the first layer of cues and the second layer of cues.
[0016] Preferably, the step S2 specifically includes the following steps:
[0017] S201, capturing directional descriptive clues in the input text T, and generating a directional bias matrix D∈R using a sigmoid function W×H×C ;
[0018] S202, the direction bias matrix D∈R W×H×C and a mask scaled to the size of the feature map The input image I is encoded by element-wise multiplication and then the CLIP visual encoder is used to extract comprehensive visual features from the encoded image.
[0019]
[0020] S203, performing a mask operation on the input image I, and then cropping it to retain only the portion containing the suggested masking area, and using the CLIP visual encoder to extract the focal visual features
[0021]
[0022] S204: Integrate the comprehensive visual features and the focal visual feature Get the comprehensive-focal visual feature F V ;
[0023]
[0024] where α is a hyperparameter between [0,1].
[0025] Preferably, the step S3 specifically includes the following steps:
[0026] S301, according to the input image I, use the pre-trained model BLIP2 to generate the title text and negative expression text, and use the CLIP text encoder to extract the title features and negative text features
[0027] S302: Based on the input text T, understand the overall meaning and target object noun phrase, and use the pre-trained CLIP text encoder to extract comprehensive text features
[0028] S303, using spatial dependency parsing to identify and select target noun phrases containing the core nouns of the sentence, and then using these selected noun phrases to extract focus text features
[0029] S304: Integrate the comprehensive text features and the focus text feature Get comprehensive-focus text features F V
[0030]
[0031] where β is a hyperparameter between [0,1].
[0032] Preferably, the step S4 specifically includes the following steps:
[0033] S401, using the pre-trained model CLIP to calculate the comprehensive-focal visual feature F V and the comprehensive-focus text feature F T The cosine similarity S V-T
[0034] S V-T = cosine(F V ,F T );
[0035] S402: Use the pre-trained model CLIP to calculate the title feature With the comprehensive-focus text feature F T The cosine similarity S C-T
[0036]
[0037] S403, using the pre-trained model CLIP to calculate the negative text features With the comprehensive-focus visual feature F V The cosine similarity S N-T
[0038]
[0039] S404, using the spatial rectifier introduced by TAS, taking the center of the mask as the position information of each mask scheme, and selecting the mask with the highest matching score in the corresponding direction area.
[0040] Preferably, the RAM++ is an open image annotation model that generates category labels for each object instance in an image through multi-granularity text supervision.
[0041] Firstly, the Grounding DINO is an open-set object detector that can accurately localize and detect object instances based on the input category label prompts, generating detailed bounding boxes for specific objects.
[0042] Preferably, the HQ-SAM is an improved version of SAM, which can generate a more detailed mask for the target area based on input points, boxes or text prompts.
[0043] Compared with the prior art, the present invention has the following beneficial effects:
[0044] 1. This paper introduces a mask generation network based on hierarchical cues, which significantly improves mask quality and reduces redundancy. In addition, a method of extracting visual features using spatial direction descriptions in textual cues is used to address the limitations of spatial sensitivity and detail recognition in the CLIP model, which will guide the CLIP visual encoder to focus on objects in a specific space in the image, significantly improving image segmentation capabilities.
[0045] 2. The present invention not only solves the problem of generating a large number of invalid and redundant masks, but also produces accurate, detailed and comprehensive instance masks even in occluded and dense scenes, showing excellent performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a schematic diagram of the method architecture of the present invention;
[0047] Figure 2 A schematic diagram of a hierarchical hint mask generation network of the present invention;
[0048] Figure 3 It is a visual display diagram of the present invention; DETAILED DESCRIPTION
[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0050] The present invention proposes a zero-shot referential image segmentation method based on hierarchical cues and directional cues, and the specific steps are as follows:
[0051] S1. Generate segmentation mask: Obtain masks of all object instances in the input image through a hierarchical hint mask generation network;
[0052] S2. Visual feature extraction and fusion: Extract comprehensive visual features based on the input image I and input text T and focal visual features At the same time, the two are fused to obtain the comprehensive-focal visual feature F V ;
[0053] S3. Get the input text T, generate the title text and negative expression text using the pre-trained model BLIP2, and extract the title features using the CLIP text encoder Negative text features Then, for the input text T, the CLIP text encoder is used to extract comprehensive text features And the focus text features At the same time, the two are combined to obtain the comprehensive-focus text feature F T ;
[0054] S4. Text prompt assisted mask selection: Use the pre-trained model CLIP to calculate the cosine similarity between the input image I and the input text T, introduce a spatial rectifier to use the center of the mask as the position information of each mask scheme, and select the mask with the highest matching score in the corresponding direction area.
[0055] Specifically, step S1 includes the following steps:
[0056] (1) Specifically, given an input image I, the network initially utilizes RAM++ to generate category labels for all object instances as the first-level cues.
[0057] (2) These labels are then used by Grounding DINO to generate object detection boxes as the second layer of cues.
[0058] (3) Finally, HQ-SAM utilizes these bounding box cues to generate segmentation masks for all object instances in the image.
[0059] Specifically, step S2 includes the following steps:
[0060] (1) Given an input text T, we capture directional descriptive cues in the text and generate a directional bias matrix using the sigmoid function.
[0061] (2) The direction bias matrix D∈R W×H×C and a mask scaled to the size of the feature map Encode the input image I by element-wise multiplication and then use the CLIP visual encoder to extract comprehensive visual features from the encoded image
[0062]
[0063] (3) In order to obtain the visual features of a specific area, the input image I is masked and then cropped (Crop()) to retain only the part containing the proposed masked area. The CLIP visual encoder is used to extract the focal visual features.
[0064]
[0065] (4) Integrate the comprehensive visual features and the focal visual features to obtain the comprehensive-focal visual feature F V .
[0066]
[0067] In fact, α is a hyperparameter between [0,1].
[0068] Specifically, step S3 includes the following steps:
[0069] (1) Given an input image I, use the pre-trained model BLIP2 to generate title text and negative expression text, and use the CLIP text encoder to extract title features and negative text features
[0070] (2) Given an input expression T, it is crucial to understand the overall meaning and the target object noun phrase, and use the pre-trained CLIP text encoder to extract comprehensive text features
[0071] (3) Use spatial dependency parsing to identify and select target noun phrases containing the core nouns of the sentence, and then use these selected noun phrases to extract focus text features
[0072] (4) Integrate comprehensive text features and focus text features to obtain comprehensive-focus text features F V .
[0073]
[0074] where β is a hyperparameter between [0,1].
[0075] Specifically, step S4 includes the following steps:
[0076] (1) Use the pre-trained model CLIP to calculate the comprehensive-focal visual feature F V and comprehensive-focus text feature F T The cosine similarity S V-T.
[0077] S V-T = cosine(F V ,F T )
[0078] (2) Use the pre-trained model CLIP to calculate title features and comprehensive-focus text feature F T The cosine similarity S C-T .
[0079]
[0080] (3) Use the pre-trained model CLIP to calculate negative text features With comprehensive-focus visual features F V The cosine similarity S N-T .
[0081]
[0082] (4) Using the spatial rectifier introduced by TAS, the center of the mask is used as the position information of each mask scheme, and the mask with the highest matching score in the corresponding direction area is selected. The final mask It is selected by choosing the one with the highest matching score S.
[0083] S=S V-T +ηS C-T +θS N-V
[0084]
[0085] The present invention also provides an innovative framework consisting of four parts: a mask generation network based on hierarchical cues, visual feature extraction fusion, text feature extraction fusion, and text clues-assisted mask selection. The mask generation network accurately generates masks based on hierarchical cues; the directional descriptions in the text are used as clues to guide the visual encoder to focus on specific areas of the image to extract comprehensive visual features and focal visual features. In addition, the text encoder is used to extract comprehensive text features and focal text features. Finally, with the help of text clues, the most relevant masks are selected by calculating the cosine similarity between all mask images and the reference expression to achieve accurate image segmentation.
[0086] The Hierarchical Cue-Based Mask Generation Network integrates three advanced models: Recognize Anything Plus Model (RAM++), Grounding DINO, and Segment Anything in High Quality (HQ-SAM). RAM++ is an open-set image annotation model that generates category labels for each object instance in an image with multi-granular text supervision. Grounding DINO, as an open-set object detector, can accurately localize and detect object instances based on the input category label cues, generating detailed bounding boxes for specific objects. HQ-SAM is an improved version of SAM, which not only improves the quality of segmentation masks, but also generates more detailed masks for target regions based on input points, boxes, or text cues. In addition, it can generate masks for all existing instances in an image without additional cues. The network initially utilizes RAM++ to generate category labels for all object instances as the first layer of cues. These labels are then used by Grounding DINO to generate object detection boxes as the second layer of cues. Finally, HQ-SAM uses these bounding box cues to generate segmentation masks for all object instances in the image. This hierarchical cue-based approach not only solves the problem of generating a large number of invalid and redundant masks, but also produces precise, detailed and comprehensive instance masks, demonstrating superior performance even in occluded and dense scenes.
[0087] Figure 2 a shows a qualitative comparison between the proposed method and the baseline method, which shows that the proposed method achieves higher accuracy in mask generation. Figure 2 b describes the hierarchical hint mask generation network, which initially utilizes RAM++ to generate category labels for all object instances as the first layer of hints. These labels are then used by Grounding DINO to generate object detection boxes as the second layer of hints. Finally, HQ-SAM utilizes these bounding box hints to generate segmentation masks for all object instances in the image. This hierarchical hint-based approach not only solves the problem of generating a large number of invalid and redundant masks, but also produces accurate, detailed and comprehensive instance masks even in occluded and dense scenes, showing excellent performance.
[0088] Since CLIP is insensitive to spatial position information, this information is inevitably reduced during the image encoding process of the model. However, referential expressions often contain clear directional clues (such as "the woman on the right", "the bread on the table", etc.), which are valuable but easily overlooked. In order to emphasize this type of position information, the present invention captures directional descriptive clues in the text and proposes a directional bias matrix. This matrix is then used to encode the image through element-by-element multiplication.
[0089] I'=I⊙D
[0090] Where I is a given image, D∈R W×H×C is the direction bias matrix, and ⊙ is the Hadamard product operation. If, after grammatical analysis, the text contains clear directional cues (such as up, down, left, and right), then D will be a direction bias matrix generated by a sigmoid function, with bias values ranging from 1 to 0 along a given direction axis. The horizontal axis represents directions such as left, right, west, and east, and the vertical axis represents directions such as up, down, north, and south. The diagonal axis represents directions such as southeast and northwest. Lower values indicate areas that should be paid less attention. If no directional cues are detected, the matrix D will be filled with 1.
[0091] Visual features are extracted from the encoded image using the CLIP pre-trained model. In order to perform a more detailed analysis of the masked region and its surrounding information, the CLIP visual encoder is optimized according to the method proposed by Global-Local CLIP. In the ResNet-50 model, the global average pooling layer is replaced by an attention pooling layer that utilizes a multi-head attention mechanism. In the ViT-B / 32 model, the image is segmented and processed through the Transformer layer, and the mask token is applied only in the final layer. The output of the class label (CLS) from ViT is used as a comprehensive visual representation.
[0092] Therefore, the comprehensive visual features of the present invention can be expressed as:
[0093]
[0094] in, To zoom to I ′ The resized mask of the feature map size.
[0095] Focused visual features. To obtain the visual features of a specific region, the image is masked and then cropped to retain only the portion containing the proposed masked region. The cropped and masked image is fed into the CLIP visual encoder, which aims to capture the focused visual features.
[0096] F f V =∮ CLIP (Crop(I⊙M))
[0097] Crop() represents the cropping operation, the purpose of which is to focus only on the target object itself.
[0098] Once the comprehensive and focused visual features are obtained, they are fused by direct addition:
[0099]
[0100] In fact, α is a hyperparameter between [0,1].
[0101] As mentioned in TAS, the domain gap between natural images and mask images affects the alignment of visual and text features, and the presence of a large number of objects in the image that are irrelevant to the referent expression can distract attention. Therefore, the present invention uses the CLIP text encoder to extract title features from the title text. Extracting negative text features from negative expressions
[0102] Comprehensive text features. Similar to visual features, it is crucial to understand the overall meaning and target object noun phrases in a given expression T. To this end, the present invention uses a pre-trained CLIP text encoder to extract comprehensive text features.
[0103] Focused text features. Although the CLIP text encoder is able to extract text features aligned with the image, the complexity is that the presence of multiple clauses makes it difficult to focus on the key nouns. To solve this problem, the present invention uses spatial dependency parsing to identify and select target noun phrases containing the core nouns of the sentence. These selected noun phrases are then used to extract focused text features.
[0104] Similar to visual features, comprehensive text features and focus text features are fused by direct addition:
[0105]
[0106] In order to identify the most relevant mask associated with the reference expression, the present invention uses the pre-trained model CLIP to calculate the cosine similarity between the mask image and the reference expression. Considering the cosine similarity of the comprehensive-focus visual text features, the title features and the comprehensive-focus text features, the negative text expression features and the comprehensive-focus visual features, the matching score is calculated. The calculation formula of the matching score is as follows:
[0107] S V-T = cosine(F V ,F T )
[0108]
[0109]
[0110] T nis a set of noun phrases. Since all scores are cosine similarities calculated in the common CLIP feature space, the linear combination of the above three scores can obtain the final matching score. As mentioned above, CLIP is insensitive to spatial relationships. Therefore, the present invention adopts the spatial rectifier introduced by TAS, takes the center of the mask as the position information of each mask scheme, and selects the mask with the highest matching score in the corresponding directional area. In this way, the present invention forces CLIP to focus on a specific area when processing directional descriptions, thereby correcting incorrect predictions.
[0111] S=S V-T +ηS C-T +θS N-V
[0112]
[0113] Final Mask It is selected by choosing the one with the highest matching score S.
[0114] Qualitative analysis. Figure 3 The qualitative comparison of the proposed method with previous methods is shown. The results in the figure show that the proposed method can effectively extract detailed features from images, understand complex reference expressions including spatial descriptors, and select the most accurate target object mask.
[0115] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.
Claims
1. A zero-shot referential image segmentation method based on hierarchical cues and directional cues, characterized in that: The following steps are involved: S1, obtain masks of all object instances in the input image through a hierarchical hint mask generation network; S2. Extract comprehensive visual features based on input image I and input text T and focal visual features At the same time, the two are fused to obtain the comprehensive-focal visual feature F V ; S3. Get the input text T, generate the title text and negative expression text using the pre-trained model BLIP2, and extract the title features using the CLIP text encoder Negative text features Then, for the input text T, the CLIP text encoder is used to extract comprehensive text features And the focus text features At the same time, the two are combined to obtain the comprehensive-focus text feature F T ; S4. Use the pre-trained model CLIP to calculate the cosine similarity between the input image I and the input text T, introduce a spatial rectifier to use the center of the mask as the position information of each mask scheme, and select the mask with the highest matching score in the corresponding direction area.
2. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 1, characterized in that: The hierarchical prompt mask generation network integrates three advanced models: Recognize Anything Plus Model (RAM++), Grounding DINO and Segment Anything in High Quality (HQ-SAM).
3. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 2, characterized in that: The step S1 specifically includes the following steps: S101, generating category labels for all object instances using RAM++ according to the input image I as the first layer of prompts; S102, using Grounding DINO to generate a target detection frame from the category label as a second-level prompt; S103 . HQ-SAM generates segmentation masks for all object instances in the image using the first layer of cues and the second layer of cues.
4. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 3, characterized in that: The step S2 specifically includes the following steps: S201, capturing directional descriptive clues in the input text T, and generating a directional bias matrix D∈R using a sigmoid function W×H×C ; S202, the direction bias matrix D∈R W×H×C and a mask scaled to the size of the feature map The input image I is encoded by element-wise multiplication and then the CLIP visual encoder is used to extract comprehensive visual features from the encoded image. S203, performing a mask operation on the input image I, and then cropping it, retaining only the portion containing the suggested masking area, and using the CLIP visual encoder to extract the focal visual features S204: Integrate the comprehensive visual features and the focal visual feature Get the comprehensive-focal visual feature F V ; where α is a hyperparameter between [0,1].
5. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 4, characterized in that: The step S3 specifically comprises the following steps: S301, according to the input image I, use the pre-trained model BLIP2 to generate the title text and negative expression text, and use the CLIP text encoder to extract the title features and negative text features S302: Based on the input text T, understand the overall meaning and target object noun phrase, and use the pre-trained CLIP text encoder to extract comprehensive text features S303, using spatial dependency parsing to identify and select target noun phrases containing the core nouns of the sentence, and then using these selected noun phrases to extract focus text features S304: Integrate the comprehensive text features and the focus text feature Get comprehensive-focus text features F V where β is a hyperparameter between [0,1].
6. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 5, characterized in that: The step S4 specifically comprises the following steps: S401, using the pre-trained model CLIP to calculate the comprehensive-focal visual feature F V and the comprehensive-focus text feature F T The cosine similarity S V-T S V-T =cosine(F V ,F T ); S402: Use the pre-trained model CLIP to calculate the title feature With the comprehensive-focus text feature F T The cosine similarity S C-T S403, using the pre-trained model CLIP to calculate the negative text features With the comprehensive-focus visual feature F V The cosine similarity S N-T S404, using the spatial rectifier introduced by TAS, taking the center of the mask as the position information of each mask scheme, and selecting the mask with the highest matching score in the corresponding direction area.
7. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 2, characterized in that: The RAM++ is an open-source image annotation model that generates category labels for each object instance in an image through multi-granular text supervision.
8. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 2, characterized in that: The Grounding DINO is an open-set object detector that can accurately localize and detect object instances based on input category label cues, generating detailed bounding boxes for specific objects.
9. The zero-shot referential image segmentation method based on hierarchical cues and directional cues according to claim 2, characterized in that: The proposed HQ-SAM is an improved version of SAM that can generate more detailed masks for target regions based on input points, boxes, or text cues.
Citation Information
Patent Citations
Segmentation recognition and text description method and system based on promptable segmentation model
CN117671688A
Image processing method, electronic device, storage medium and computer program product
CN118229602A
Image processing method, apparatus and device, and storage medium
WO2020216008A1
Cited By
Zero sample object instance segmentation system and method in robot environment
CN119784778A
Zero sample sketch retrieval method based on mask and matching
CN119848283A
A zero-shot sketch retrieval method based on mask and matching
CN119848283B
Medical ultrasound image data set construction method and system and storage medium
CN120766056A
Zero sample anaphora image segmentation method based on text perception adapter
CN121725007A
