Method and system for enhancing emotion detection by using generative image semantics
By generating descriptive text from generative images and fusing it with the original text, the semantic alignment problem between image and text modalities is solved, thereby improving the interpretability and robustness of sentiment detection. This approach is applicable to multimodal sentiment detection scenarios such as social media.
Patent Information
- Application Number
- CN202511121568.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-12
- Publication Date
- 2025-11-18
AI Technical Summary
Existing sentiment detection methods lack deep semantic alignment between image and text modalities. Image semantics are mostly in the form of embedded vectors, which lack interpretability and are difficult to assist text in supplementing emotions. Furthermore, generative large models have not been effectively utilized in sentiment detection tasks.
Generative images generate descriptive text, transforming images into descriptive text with natural language capabilities. This text is then fused with the original text to form enhanced text. A pre-trained language model is used to extract multimodal semantic representation vectors, which are then input into a sentiment classification model for judgment.
It improves the interpretability of image emotion understanding and the synergy of image-text semantic fusion, enhances the robustness and generalization ability of emotion detection, and is suitable for emotion recognition in complex, ambiguous or low-resource scenarios.
Smart Images
Figure CN120974236A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology and relates to a method and system for emotion detection, particularly a method and system for emotion detection using generative image semantic enhancement. Background Technology
[0002] With the widespread use of social media, short video platforms, and content generation tools, users' emotional expression is increasingly showing complex characteristics such as text-image hybridization, multimodal fusion, and semantic ambiguity. Images not only serve as a supplementary medium to text, but also often carry deeper information such as emotional cues, stances and attitudes, and cultural satire, becoming an important data source in sentiment analysis.
[0003] In recent years, research has attempted to incorporate image modalities into emotion detection tasks to compensate for the limitations of text features. These methods can be broadly categorized as follows: (1) Image feature-assisted fusion method.
[0004] These methods typically extract static visual features using image encoders (such as ResNet, VGG, CLIP-Vision, etc.), and then concatenate or fuse these features with textual features before feeding them into the sentiment classification model. Some studies also integrate visual attributes such as color saturation, subject position, and image style to aid in judging emotional tendencies. These methods improve the model's sensitivity to image context to some extent, but because image semantics are represented only in the form of vector embeddings, they lack clear emotional interpretation capabilities and differ significantly from the semantic space of text, resulting in limited fusion effects and insufficient interpretability of sentiment classification models.
[0005] (2) Image-text sentiment consistency modeling method.
[0006] These methods aid sentiment inference by measuring the semantic consistency between images and text. The main idea is that if the sentiment of the image and text are consistent, it is more likely to be a genuine expression; conversely, if they are inconsistent, it may be a masking of emotion or hypothetical content. Common methods include using cross-modal models such as CLIP to calculate image-text similarity, image-sentiment tag alignment, and contrastive learning. Although these methods have made some progress in image-text co-modeling, they still rely on black-box similarity calculations, lacking structured modeling of the emotional components of images and failing to explicitly explain how images participate in sentiment judgment.
[0007] (3) Image emotion classification or facial expression recognition methods.
[0008] These methods treat images as independent channels for emotion judgment, using existing emotion classification models (such as ImageEmotion, FacialEmotionCNN, etc.) to directly classify the emotions in images, or combining them with facial expression analysis tools to detect basic emotions such as joy, anger, sorrow, and happiness. These methods perform well under conditions of standardized images or rich facial features, but social media platforms commonly suffer from unstructured images, complex scenes, and blurry, low-quality images, leading to unstable model performance and insufficient generalization ability. Furthermore, these methods struggle to semantically coordinate with text, providing only local cues and limiting the overall accuracy of emotion recognition.
[0009] Therefore, although the above methods have expanded the modal boundaries of sentiment detection to some extent, the following key problems still need to be solved: (1) Image semantics are mostly in the form of embedded vectors, lacking interpretable language expression, and are difficult to assist text in supplementing emotions; (2) Most image-text fusion methods remain at the feature level or similarity level, and cannot achieve deep semantic alignment; (3) Generative large models have not yet been systematically applied to image-text sentiment collaborative modeling, and their image generation capabilities have not been effectively utilized by sentiment detection tasks.
[0010] Given the aforementioned shortcomings of existing technologies, there is an urgent need to research a new emotion detection method and system. Summary of the Invention
[0011] To overcome the shortcomings of existing technologies, this invention proposes a method and system for emotion detection using generative image semantic enhancement, which significantly improves the interpretability of image emotion understanding, the synergy of image-text semantic fusion, and the system's deployment efficiency and generalization ability.
[0012] To achieve the above objectives, the present invention provides the following technical solution: A method for emotion detection using generative image semantic enhancement, characterized by the following steps: 1) Obtain tweet data containing text content and accompanying images from social media, and preprocess the text content and accompanying images; 2) Based on image-to-text technology, the preprocessed accompanying images are converted into descriptive text with natural language expression capabilities; 3) The preprocessed text content is merged with the descriptive text to form the merged enhanced text; 4) Input the fused enhanced text into a pre-trained language model to extract text feature codes, use a visual encoder to encode the pre-processed matching images to extract image feature codes, and jointly model the text feature codes and image feature codes to generate a unified multimodal semantic representation vector; 5) Input the multimodal semantic representation vector into the sentiment classification model, and let the sentiment classification model determine the sentiment that the tweet data wants to express.
[0013] Preferably, in step 1), preprocessing the text content specifically involves: performing a normalization operation on the text content to remove noise, including removing HTML tags, special characters, URL links, and meaningless punctuation.
[0014] Preferably, in step 1), the preprocessing of the accompanying images specifically involves: standardizing the size of the accompanying images and removing images that are missing or blurry.
[0015] Preferably, step 2) specifically involves: using a pre-trained multimodal generation model to convert the accompanying images into descriptive text with natural language expression capabilities; and when the tweet data contains multiple accompanying images, using the pre-trained multimodal generation model to convert each accompanying image into a descriptive text with natural language expression capabilities; and then using a large language model to uniformly summarize and abstract the descriptive texts corresponding to each accompanying image to generate a graphic summary.
[0016] Preferably, in step 3), the preprocessed text content is merged with the descriptive text and the graphic summary content to form the merged enhanced text.
[0017] Preferably, in step 3), the fusion method used during fusion includes prefix splicing fusion, suffix splicing fusion, or structured template fusion.
[0018] Preferably, in step 5), the emotion classification model uses a discriminant head composed of a multilayer perceptron.
[0019] Furthermore, the present invention also provides a system for enhancing sentiment detection using generative image semantics, characterized in that it includes: A multimodal data acquisition and preprocessing module is used to acquire tweet data containing text content and accompanying images from social media, and to preprocess the text content and accompanying images. The image-to-text module is used to convert the accompanying images into descriptive text with natural language expression capabilities based on image-to-text technology. The image-text fusion enhancement module is used to fuse the preprocessed text content with the descriptive text to form the fused enhanced text; The image and text representation module is used to input the fused enhanced text into a pre-trained language model to extract text feature codes, use a visual encoder to encode the pre-processed accompanying image to extract image feature codes, and jointly model the text feature codes and image feature codes to generate a unified multimodal semantic representation vector. The sentiment classification module is used to input the multimodal semantic representation vector into the sentiment classification model, which then determines the sentiment that the tweet data wants to express.
[0020] Furthermore, the present invention also provides a generative image semantic enhancement emotion detection device, characterized in that it includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the generative image semantic enhancement sentiment detection method as described above. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the steps of the generative image semantic enhancement emotion detection method as described above.
[0021] Compared with existing multimodal sentiment detection methods that rely on "image feature stitching + image-text similarity scoring," the generative image semantic enhancement sentiment detection method and system proposed in this invention have the following significant advantages in terms of interpretability of image sentiment understanding, synergy of image-text semantic fusion, and system deployment efficiency and generalization ability: (1) Expressing emotions in images in a language-based manner, improving semantic alignment and interpretability.
[0022] This invention innovatively introduces a generative image description mechanism, converting the original image (and accompanying pictures) into a text description with natural language structure through a pre-trained image-text generation model (such as BLIP, GPT-4V, etc.), thereby achieving a structured linguistic expression of image emotion, scene, and behavioral information. This approach overcomes the limitations of traditional image feature embedding and text semantic inconsistencies, enabling the fusion of image and text information in a unified semantic space, and improving the interpretability of image emotional information through natural language description.
[0023] (2) Semantic enhancement fusion strategy to improve emotion understanding ability.
[0024] This invention combines multi-dimensional descriptive text generated from images with original user text (text content) to construct an enhanced input corpus, integrating multimodal emotional information through language as a medium. Compared to traditional feature concatenation methods, this approach offers stronger contextual coherence and consistency in emotional expression, helping emotion recognition models (such as RoBERTa and BERT) accurately capture emotional signals in image-text combinations.
[0025] (3) It is robust and suitable for scenarios with imbalance of text and image information and fluctuation of image quality.
[0026] This invention employs an image semantic textification strategy, effectively mitigating common social media issues such as inconsistencies between images and text, blurry images, or images that have been tampered with. Even with low-quality images or a lack of explicit tags, it can still extract latent semantic clues to aid in sentiment assessment, demonstrating stronger task robustness and cross-content generalization capabilities. Attached Figure Description
[0027] Figure 1 This is a flowchart of the generative image semantic enhancement emotion detection method of the present invention.
[0028] Figure 2 This is a schematic diagram of the generative image semantic enhancement emotion detection system of the present invention. Detailed Implementation
[0029] Before detailing any embodiment of the invention, it should be understood that the invention, in its application, is not limited to the details of the construction and arrangement of the components set forth in the following description or illustrated in the following figures. The invention can have other embodiments and can be practiced or carried out in various ways. Furthermore, it should be understood that the wording and terminology used herein are for descriptive purposes and should not be considered limiting. The use of “comprising” or “having” and variations thereof is intended to cover the items set forth below and their equivalents, as well as any additional items. Unless otherwise specified or limited, the terms “installation,” “connection,” “support,” and “linkage,” and variations thereof are used broadly and cover both direct and indirect installation, connection, support, and linking. Moreover, “connection” and “linkage” are not limited to physical or mechanical connections or links. Furthermore, firstly, in the disclosure of this invention, the terms "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the above terms should not be construed as limiting this invention. Secondly, the term "a" should be understood as "at least one" or "one or more," that is, in one embodiment, the number of an element can be one, while in another embodiment, the number of the element can be multiple. The term "a" should not be construed as a limitation on the quantity.
[0030] With the surge in user-generated content (UGC) in applications such as social media, short video platforms, and intelligent customer service, sentiment detection, as a crucial technology for understanding user attitudes and emotions, has been widely applied in areas such as satisfaction assessment, public opinion monitoring, and product evaluation. Traditional sentiment detection methods primarily rely on text content for sentiment classification, making it difficult to cover multimodal scenarios such as mixed text and image data. In recent years, with the widespread availability of multimodal data, image-text sentiment detection has become a research hotspot, with mainstream methods often improving recognition performance through the fusion of image and text features. However, existing methods still face numerous challenges in practical applications: First, images and text belong to different modalities with significantly different semantic expressions. Directly employing feature fusion or shallow attention mechanisms struggles to achieve deep emotional semantic alignment, especially when image and text expressions are inconsistent or ambiguous, easily leading to misjudgments. Second, social media texts are often short, subtle, and ironic, with unclear emotional expressions, making it difficult to accurately reconstruct user emotions relying solely on text. Third, traditional visual models extract mostly low-level visual features, lacking a deep understanding and structured expression of scene, behavior, and emotional features within images, resulting in the underutilization of image information's value in emotion detection. Finally, current methods lack effective semantic enhancement mechanisms and interpretability, making it difficult to maintain robustness and generalization capabilities in complex and diverse real-world scenarios. Therefore, an innovative method is urgently needed to bridge the modal semantic gap, enhance semantic understanding, and improve the accuracy and interpretability of emotion recognition.
[0031] Therefore, this invention addresses the aforementioned problems by proposing a sentiment detection method that utilizes generative image generation to enhance descriptive text. Its core idea includes: using a large-scale pre-trained generative model to verbalize image content, converting image semantics into understandable auxiliary text information; and fusing the original text with the generated text to perform semantic enhancement and multimodal sentiment modeling. This improves the accuracy, robustness, and interpretability of sentiment recognition in complex, ambiguous, or low-resource scenarios.
[0032] Figure 1 A flowchart of the generative image semantic enhancement sentiment detection method of the present invention is shown. Figure 1 As shown, the method for emotion detection using generative image semantic enhancement of the present invention includes the following steps: I. Multimodal data acquisition and preprocessing module.
[0033] We acquire tweet data, including text content and accompanying images, from social media platforms such as Weibo. For the text content, we collect the natural language content of the tweet, including the body and hashtags; for the accompanying images, we collect one or more images attached to the tweet. All tweet data is organized on a tweet-by-twee basis to ensure that the text content corresponds to the accompanying image.
[0034] In terms of text content preprocessing, the original text content undergoes normalization operations, including removing noise such as HTML tags, special characters, URL links, and meaningless punctuation. Furthermore, to ensure the semantic integrity of the text, key phrases, topic tags, and sentiment expression components need to be preserved during preprocessing. Specifically, firstly, TF-IDF, TextRank, and named entity recognition methods (such as BERT-NER) are used to extract high-semantic-weight phrases and named entities from the text content. Secondly, topic tags (such as "#movierecommendation") are extracted using regular expressions and preserved with special tags; for sentiment expression words (such as "happy," "annoying," etc.), sentiment lexicon annotation is used, and the cleaning logic is optimized to ensure that these emotionally charged words are not mistakenly deleted. Through these mechanisms, while normalizing the structure of the text content, the semantic core and sentiment cues in the user's expression are effectively preserved, providing an accurate and high-quality input foundation for subsequent processing.
[0035] Regarding the processing of accompanying images, the collected images are standardized in size, and missing or particularly blurry images are removed (generally blurry images can be retained).
[0036] By preprocessing the text content and accompanying images, we can construct image-text samples with clear structure and complete semantics, providing a unified and high-quality input for subsequent image-to-text generation and fusion enhancement.
[0037] II. Image to Text Generation.
[0038] To fully extract the semantic information contained in the accompanying images of tweets, this invention employs image-to-text technology to analyze these images. Specifically, it utilizes large-scale pre-trained multimodal generative models (such as BLIP-2, InstructBLIP, or MiniGPT-4) to convert the accompanying images into concise descriptive text with natural language capabilities.
[0039] In multimodal emotion detection tasks, especially in scenarios involving accompanying images, the quality of the Prompt design directly determines the emotional expression effect of the generated text. Prompt design needs to ensure that the generated descriptive text is accurate, emotionally oriented, concise, and possesses consistency and summarization capabilities to improve the effectiveness of multimodal emotion detection.
[0040] The Prompt used in this invention follows a three-part structure of "scene + content + emotion". For example, an exemplary Prompt would be: This is an image of [type], which shows [main content] and expresses [emotion or atmosphere].
[0041] In addition, when a tweet contains multiple accompanying images, a large-scale pre-trained multimodal generative model (such as BLIP-2, InstructBLIP, or MiniGPT-4) is first used to convert each accompanying image into a concise descriptive text with natural language expression capabilities. Then, a large language model is used to uniformly summarize and abstract the descriptive texts corresponding to each accompanying image to generate a clear and comprehensive image-text summary to improve the overall semantic density.
[0042] By generating text from images, information transfer and transformation from visual to linguistic modalities are realized, providing semantic supplementation for subsequent image-text fusion enhancement, while improving the ability to understand the collaborative context of images and text.
[0043] III. Enhanced integration of text and images.
[0044] The descriptive text and image summary generated from the accompanying images are merged with the original text content in the tweet data to construct the merged enhanced text, thereby improving the semantic expressiveness of the text.
[0045] In this invention, the fusion methods include prefix splicing, suffix splicing, or structured template fusion, which are used to simulate human graphic and text comprehension habits and maintain text coherence.
[0046] To adapt to the input requirements of downstream large language models, this invention needs to control the total length of the fused enhanced text to avoid redundant information interfering with the main semantics. When controlling the total length of the fused enhanced text, the length that can be received by subsequent pre-trained language models (such as RoBERTa) should be taken into account.
[0047] Through image-text fusion enhancement, the semantic information implicit in the accompanying images is effectively transferred to the text modality, providing support for the perception of multimodal semantics under subsequent single text input.
[0048] IV. Visual and textual representation.
[0049] The fused enhanced text is first input into a pre-trained language model (such as RoBERTa). After modeling through a multi-layer Transformer structure, a high-dimensional text representation vector containing rich contextual semantic information is extracted, i.e., text feature encoding. Simultaneously, a visual encoder (such as CLIP-ViT or ResNet) is used to encode the original accompanying image, extracting its global visual feature representation, i.e., image feature encoding. Then, the text feature encoding and image feature encoding are jointly modeled to finally generate a unified multimodal semantic representation vector. This multimodal semantic representation vector not only integrates the linguistic features of the descriptive text generated based on the accompanying image and the original text content, but also retains the visual semantic information of the accompanying image itself, providing a more comprehensive contextual understanding capability.
[0050] V. Emotional Classification.
[0051] The multimodal semantic representation vector is input into the constructed sentiment classification model (such as a discriminant head composed of a multilayer perceptron MLP) to determine the sentiment that the tweet wants to express.
[0052] The sentiment classification model can be trained through supervised learning, with the goal of minimizing the difference between the predicted results and the true labels in order to achieve effective judgment of user emotions.
[0053] Figure 2 A schematic diagram illustrating the structure of the generative image semantic enhancement emotion detection system of the present invention is shown. Figure 2 As shown, the system for generative image semantic enhancement of emotion detection according to the present invention includes: 1. Multimodal data acquisition and preprocessing module.
[0054] The multimodal data acquisition and preprocessing module is used to acquire tweet data containing text content and accompanying images from social media, and to preprocess the text content and accompanying images.
[0055] 2. Image to Text Module.
[0056] The image-to-text module is used to convert the accompanying images into descriptive text with natural language expression capabilities based on image-to-text technology.
[0057] 3. Enhanced image and text fusion module.
[0058] The image-text fusion enhancement module is used to fuse the preprocessed text content with the descriptive text to form the fused enhanced text.
[0059] 4. Graphic and text representation module.
[0060] The image-text representation module is used to input the fused enhanced text into a pre-trained language model to extract text feature codes, use a visual encoder to encode the pre-processed accompanying image to extract image feature codes, and jointly model the text feature codes and image feature codes to generate a unified multimodal semantic representation vector.
[0061] 5. Emotion Classification Module.
[0062] The sentiment classification module is used to input the multimodal semantic representation vector into the sentiment classification model, and the sentiment classification model determines the sentiment that the tweet data wants to express.
[0063] The generative image semantic enhancement sentiment detection system of this invention can be deployed in practical application environments, serving various scenarios such as public opinion monitoring systems, comment moderation platforms, and sentiment trend analysis tools. Deployment methods can include front-end application integration, back-end server deployment, or providing sentiment analysis service interfaces externally via API. The system can connect to real-time data streams from social media platforms, short video platforms, or content aggregation platforms, automatically receiving tweet data containing text content and accompanying images, sequentially completing steps such as image-to-text generation, text enhancement and fusion, image-text representation extraction, and sentiment classification, ultimately outputting the sentiment recognition result.
[0064] The generative image semantic enhancement emotion detection system of the present invention has good platform adaptability, cross-language compatibility and multimodal scalability, and can be widely used in emotion perception and management tasks of various types of social media content.
[0065] Furthermore, the present invention also provides a generative image semantic enhancement emotion detection device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the generative image semantic enhancement emotion detection method as described above. Finally, the present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the steps of the generative image semantic enhancement emotion detection method as described above.
[0066] This invention innovatively expresses image information (with accompanying pictures) in natural language and guides emotional signals to collaborate with the original text content within the text space. Building upon image features, it introduces a language enhancement path for emotional semantics, effectively improving adaptability to complex scenarios such as irony, ambiguous emotions, and multimodal conflicts. Simultaneously, this strategy enhances the ability to judge emotional consistency between images and text and to learn across modalities, making it suitable for image- and text-intensive content environments such as social media platforms and short video platforms.
[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention based on the concept of the present invention, without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A method for emotion detection using generative image semantic enhancement, characterized in that, Includes the following steps: 1) Obtain tweet data containing text content and accompanying images from social media, and preprocess the text content and accompanying images; 2) Based on image-to-text technology, the preprocessed accompanying images are converted into descriptive text with natural language expression capabilities; 3) The preprocessed text content is merged with the descriptive text to form the merged enhanced text; 4) Input the fused enhanced text into a pre-trained language model to extract text feature codes, use a visual encoder to encode the pre-processed matching images to extract image feature codes, and jointly model the text feature codes and image feature codes to generate a unified multimodal semantic representation vector; 5) Input the multimodal semantic representation vector into the sentiment classification model, and let the sentiment classification model determine the sentiment that the tweet data wants to express.
2. The method for emotion detection using generative image semantic enhancement according to claim 1, characterized in that, In step 1), the preprocessing of the text content specifically involves: performing a normalization operation on the text content to remove noise, including removing HTML tags, special characters, URL links, and meaningless punctuation.
3. The method for emotion detection using generative image semantic enhancement according to claim 1, characterized in that, In step 1), the preprocessing of the accompanying images specifically involves: standardizing the size of the accompanying images and removing images that are missing or blurry.
4. The method for emotion detection using generative image semantic enhancement according to claim 1, characterized in that, Step 2) specifically involves: using a pre-trained multimodal generation model to convert the accompanying images into descriptive text with natural language expression capabilities; and when the tweet data contains multiple accompanying images, using the pre-trained multimodal generation model to convert each accompanying image into a descriptive text with natural language expression capabilities; and then using a large language model to uniformly summarize and abstract the descriptive texts corresponding to each accompanying image to generate a text-image summary.
5. The method for emotion detection using generative image semantic enhancement according to claim 4, characterized in that, In step 3), the preprocessed text content is merged with the descriptive text and the graphic summary content to form the merged enhanced text.
6. The method for emotion detection using generative image semantic enhancement according to claim 5, characterized in that, In step 3), the fusion method used during fusion includes prefix splicing fusion, suffix splicing fusion, or structured template fusion.
7. The method for emotion detection using generative image semantic enhancement according to any one of claims 1-6, characterized in that, In step 5), the emotion classification model uses a discriminant head composed of a multilayer perceptron.
8. A system for emotion detection using generative image semantic enhancement, characterized in that, include: A multimodal data acquisition and preprocessing module is used to acquire tweet data containing text content and accompanying images from social media, and to preprocess the text content and accompanying images. The image-to-text module is used to convert the accompanying images into descriptive text with natural language expression capabilities based on image-to-text technology. The image-text fusion enhancement module is used to fuse the preprocessed text content with the descriptive text to form the fused enhanced text; The image and text representation module is used to input the fused enhanced text into a pre-trained language model to extract text feature codes, use a visual encoder to encode the pre-processed accompanying image to extract image feature codes, and jointly model the text feature codes and image feature codes to generate a unified multimodal semantic representation vector. The sentiment classification module is used to input the multimodal semantic representation vector into the sentiment classification model, and the sentiment classification model determines the sentiment that the tweet data wants to express.
9. A generative image semantic enhancement emotion detection device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the generative image semantic enhancement sentiment detection method as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the generative image semantic enhancement emotion detection method as described in any one of claims 1-7.