Art work customization generation method based on electroencephalogram emotion recognition

Abstract

The invention relates to an artistic work customization generation method based on electroencephalogram emotion recognition, and aims to drive artistic creation through emotion analysis, the method comprises two key steps of emotion recognition and work generation: firstly, obtaining a user neural signal in real time through a high-precision electroencephalogram acquisition device, and processing the user neural signal by using an emotion recognition model; the method comprises the following steps that firstly, an emotion parameter reflecting the emotion state of a user is output, then, a work generation module selects matched artistic style characteristics according to the emotion parameter, and an artistic work meeting the emotion state of the user is automatically created through a generation model. Compared with the prior art, the personalized art works can be customized according to the emotional requirements of the user.

Description

Technical Field

[0001] The present invention relates to the technical field of artwork generation, and in particular to a customized artwork generation method based on electroencephalogram emotion recognition. Background Art

[0002] With the rapid development of brain science and artificial intelligence technology, human understanding and expression of emotions are undergoing major changes. In artistic creation, emotion is a core driving force, but the traditional creative method mainly relies on the subjective emotional expression of the creator. This method is not only difficult to quantify, but also easily affected by the creator's personal emotional fluctuations, thus causing instability in the quality of the artwork and the effect of emotional transmission. In addition, the viewer's emotional interpretation of the artwork often deviates from the original intention of the creator, resulting in inconsistency in emotional transmission. Therefore, how to use technical means to customize the complex and diverse emotional states of human beings into artworks with distinct emotional expressions, and realize the full automation of this process, has become an important research direction in the current multidisciplinary cross-field.

[0003] In recent years, the application of EEG signals in emotion recognition has made some progress. For example, by analyzing EEG signals such as alpha waves and beta waves, the user's basic emotional state (such as happiness, sadness, anger, etc.) can be identified, and attempts can be made to transform these emotions into art forms such as music and painting. Some studies have introduced artificial intelligence technologies, such as generative adversarial networks (GANs) and variational autoencoders (VAE), to achieve the mapping of emotions to artistic expressions. However, these methods usually only focus on certain fixed emotional states, lack support for complex and diverse emotional needs, and have weak detail expression and personalized expression capabilities in the process of generating artworks. In addition, existing technologies have failed to achieve automation of the entire process from emotion recognition to art generation, relying on manual intervention, and are difficult to meet the actual needs of efficient creation and real-time emotional expression.

[0004] Although the application of EEG signals in emotion recognition and artistic creation has shown broad prospects, it still faces two core challenges in practical applications. First, the accuracy of emotion recognition is significantly affected by factors such as individual differences and environmental interference, and it is difficult to stably capture the complex emotional states of dynamic changes, which limits the reliability and universality of emotional signals. Second, the mapping mechanism between emotions and artistic styles lacks deep modeling. Current technology is difficult to meet the needs of customized expression of different emotional states, and the generated works are insufficient in diversity, expressiveness and personalization. At the same time, the overall process from EEG emotion recognition to the generation of artworks has not yet been fully automated. The reliance on manual participation is not only inefficient, but also limits the application potential of technology in real scenes. These problems show that there is an urgent need to build an art generation technology that can accurately identify diverse emotions, customize the expression of different emotional states, and realize full-process automation to meet the actual needs of personalized emotional expression and efficient artistic creation. Summary of the invention

[0005] The purpose of the present invention is to provide a customized generation method of artworks based on EEG emotion recognition, which can accurately identify the user's emotional state and map it to a corresponding artistic style in a customized manner, while realizing the full-process automated generation of artworks.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A method for customized generation of artworks based on EEG emotion recognition, comprising the following steps:

[0008] S1, by collecting the user's neural signals, extracting feature data related to the emotional state, and converting the feature data into emotional parameters representing the intensity and pleasure of emotion through a quantitative model;

[0009] S2, dynamically matches the target artistic style characteristics based on the emotional parameters, and uses the generative model to automatically generate artworks.

[0010] The S1 comprises the following steps:

[0011] S101, collecting user emotional features: obtaining the user's immediate emotional state through neural signals to form a multi-dimensional feature matrix;

[0012] S102, quantification and labeling of emotional state: based on the constructed emotional evaluation system, quantify the user's emotional experience, obtain emotional parameters, label the multi-dimensional feature matrix, and form an emotional feature data set;

[0013] S103, preprocessing of emotional data: cleaning, denoising and normalizing the emotional feature data set;

[0014] S104, construction and training of emotion recognition model: construct an emotion recognition model using machine learning algorithm, train the emotion recognition model based on the emotion feature dataset, and learn the mapping relationship between neural signals and emotion parameters;

[0015] S105, output of emotional state: based on the emotion recognition model, generate real-time quantitative parameters of the user's emotional state.

[0016] The neural signal includes an electroencephalogram signal, specifically, a power spectrum density value in different frequency bands.

[0017] The quantification of the user's emotional experience is specifically as follows: converting the user's emotional state into an Arousal value and a Valence value, wherein the Arousal value represents the degree of arousal and the Valence value represents the degree of pleasure.

[0018] The S2 comprises the following steps:

[0019] S201, emotion input and style extraction: dynamically select the target art style features based on the emotion parameters and extract the core elements of the artwork generation;

[0020] S202, generative model loading and fine-tuning: quickly fine-tune the pre-trained generative model through low-rank adaptation technology to adapt it to the target artistic style;

[0021] S203, Emotion and style mapping optimization: Based on the transfer learning technology, the generation model's ability to generate diverse artistic styles is expanded, and a mapping mechanism from emotion parameters to diverse styles is constructed;

[0022] S204, Automated Generation and Quality Optimization: Use the generative model to automatically generate artworks that meet the emotional parameters, and improve the quality and consistency of the generated artworks through optimization algorithms.

[0023] The dynamic selection of the target art style features includes extracting key color features, brushstroke features and composition layout features in the artwork, and the style features are adaptively selected according to the user's emotional intensity and emotional tendency, wherein the Arousal value and Valence value are divided into two levels, high and low, according to a preset threshold.

[0024] When both the Arousal value and the Valence value are high, choose bright colors and strong color contrast, use more curves and dynamic composition, and express a highly active and pleasant emotional state; when the Arousal value is high and the Valence value is low, choose cool colors, dark tones, and a messy composition, reflecting an intense but negative emotional tendency; when the Arousal value is low and the Valence value is high, use soft tones, keep the composition symmetrical and the elements sparse, to express tranquility and pleasant emotions; when both the Arousal value and the Valence value are low, use a dark background and low-saturation colors, with a simple straight line composition, to convey a low and negative emotional experience.

[0025] In the generative model loading and fine-tuning, the pre-trained GLM-4 large language model is used, and the low-rank adaptation technology is used to insert a low-rank matrix into the Transformer layer of the pre-trained model to perform model fine-tuning.

[0026] Specifically, S203 is as follows: model training is performed based on the existing Chinese heavy-color traditional Chinese painting database, and the model's ability to adapt to artistic styles is expanded through transfer learning. During the transfer learning process, the original traditional Chinese painting data is replaced with works of various artistic styles including oil painting and watercolor painting. The same process as the traditional Chinese painting model training is adopted, including the use of Stable Diffusion technology, combined with low-rank adaptation and LoCon algorithms, to extract artistic style features.

[0027] In S204, a real-time feedback mechanism is introduced to adjust the emotional parameters according to the user's feedback on the generated work, thereby achieving optimization and adjustment of the generated art work.

[0028] The artworks include digital paintings and virtual reality art.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) Accurate and efficient emotion recognition: This invention uses EEG signals to perform emotion recognition, which can accurately capture the complex emotional state of an individual, breaking through the problem of difficulty in quantifying emotional expression in traditional artistic creation. By dynamically analyzing emotional dimensions such as arousal and pleasure, a deep mapping of emotion to artistic style is achieved.

[0031] (2) Customization and diversity: This paper combines low-rank adaptation (LoRA) and transfer learning technology to support the rapid generation of a variety of artistic styles, covering everything from traditional Chinese painting to modern painting styles. Through emotion-driven customized generation of artistic styles, it meets the user's personalized emotional expression needs.

[0032] (3) Full-process automated design: This invention realizes the full-process automated operation of emotion recognition, style selection and art generation without manual intervention, greatly improving the creation efficiency. At the same time, through efficient model fine-tuning technology, the computing cost is significantly reduced, and multi-scenario deployment and real-time response are supported. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of the method of the present invention;

[0034] Figure 2 A diagram of a relationship between emotion and style mapping provided in an embodiment of the present invention;

[0035] Figure 3 This is a customized artwork result diagram under different emotions provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0037] It should be pointed out in particular that the quantification parameters of the emotional state (such as Arousal value and Valence value), the extraction method of artistic style, and the low-rank adaptation technology (LoRA) mentioned in the present invention are all preferred implementation methods. Its purpose is to accurately identify the emotional state of the user and generate artworks that meet the needs based on the emotional input. Although the embodiments of the present invention use these specific methods, the core idea of ​​the present invention is not limited to the use of these specific quantification methods and techniques. Within the scope of protection of the present invention, any technical solution that can achieve an effective connection between emotional state and art generation, and can generate personalized artworks according to user needs, can be regarded as the protection content of the present invention.

[0038] Example 1

[0039] This embodiment provides a method for customizing artwork generation based on EEG emotion recognition. Figure 1 As shown, the following steps are included:

[0040] S1, by collecting the user's neural signals, extracts feature data related to the emotional state, and converts the feature data into emotional parameters that represent the intensity and pleasure of emotion through a quantitative model.

[0041] S1 includes the following steps:

[0042] S101, collection of user emotional features: obtaining the user's immediate emotional state through neural signals to form a multi-dimensional feature matrix.

[0043] Neural signals include EEG signals, specifically the power spectral density values ​​of different frequency bands such as alpha waves and beta waves.

[0044] "Neural signals" specifically refer to brain electrical activity signals collected by electroencephalogram (EEG) equipment. EEG detects and records the potential changes caused by the activity of neurons in the cerebral cortex by attaching electrodes to the scalp. The signal is usually manifested as waves in different frequency bands (such as Delta, Theta, Alpha, Beta, etc.). EEG signals can provide dynamic information about brain activity, which is especially valuable for identifying emotional states.

[0045] In the specific collection process, high-precision EEG instruments (such as Emotiv or NeuroSky) are used to capture the electrical activity of the cerebral cortex in real time when it perceives stimulation (such as watching pictures, videos or auditory stimulation). By recording these signals, the user's immediate emotional state can be reflected as input data.

[0046] S102, quantification and labeling of emotional state: Based on the constructed emotional evaluation system, quantify the user's emotional experience, obtain emotional parameters, label the multi-dimensional feature matrix, and form an emotional feature data set.

[0047] This embodiment converts the user's emotional state into an Arousal value and a Valence value, wherein the Arousal value represents the degree of arousal and the Valence value represents the degree of pleasure. These parameters represent the user's emotional intensity and the positive or negative tendency of the emotion, respectively.

[0048] The core of the emotion quantification method is to use a standardized emotion evaluation scale (such as the SAM scale), where the Arousal value ranges from -1 to 1, indicating a very low arousal (relaxation) to a very high arousal (excitement); the Valence value ranges from -1 to 1, indicating an extremely negative emotional experience (such as anger or sadness) to an extremely positive emotional experience (such as happiness or excitement). Quantified emotional data can provide more accurate and personalized emotional input for the generation of artworks.

[0049] S103, emotional data preprocessing: cleaning, denoising and normalizing the emotional feature data set.

[0050] In order to improve the accuracy of the emotion recognition model, a data preprocessing step is used before the data is input into the model. First, the collected EEG signals are cleaned to remove redundant and noisy data. Then, the data is normalized to ensure that the data values ​​fall within a uniform scale, which is convenient for subsequent processing and training of the machine learning model.

[0051] S104, Construction and training of emotion recognition model: Use machine learning algorithm to build emotion recognition model, train the emotion recognition model based on emotion feature dataset, and learn the mapping relationship between neural signals and emotion parameters.

[0052] The core algorithm of emotion recognition is the machine learning model. In this process, the random forest regressor is used, which is a powerful supervised learning algorithm that is suitable for processing high-dimensional feature data and can effectively map the relationship between the input EEG features and the target emotion value. During the model training process, a large amount of labeled data (including EEG signals and corresponding emotion scores) is used, and the model performance is optimized through parameter adjustment and cross-validation to improve the accuracy of emotion quantification.

[0053] S105, output of emotional state: based on the emotion recognition model, generate real-time quantitative parameters of the user's emotional state.

[0054] The trained emotion recognition model can analyze the EEG data collected by the user in real time and output the corresponding emotional state quantitative parameters (such as Arousal value and Valence value). These emotional parameters provide key customized inputs for the generation of artworks in step S2, ensuring that the generation of artworks can closely meet the emotional needs of users.

[0055] S2, dynamically matches the target artistic style characteristics based on the emotional parameters, and uses the generative model to automatically generate artworks.

[0056] S2 includes the following steps:

[0057] S201, emotion input and style extraction: dynamically select the target art style features based on the emotion parameters and extract the core elements of the artwork generation.

[0058] Dynamic selection of target art style features includes the extraction of key color features, brushstroke features, and composition layout features in artworks. Style features are adaptively selected based on the user's emotional intensity and emotional tendency. The style features corresponding to each emotional state (such as high Arousal positive emotion) need to be accurately mapped in the art generation process. For example, when the emotional parameters indicate that the user's pleasure is high, warm tones and bright colors will be selected to generate artworks.

[0059] S202, generative model loading and fine-tuning: The pre-trained generative model is quickly fine-tuned through low-rank adaptation technology (LoRA) to adapt it to the target artistic style.

[0060] This example uses the pre-trained GLM-4 large language model. These models have been pre-trained on a large art dataset and can learn the basic features of various art styles. On this basis, the model is fine-tuned using the low-rank adaptation technique (LoRA), so that the model can be flexibly adjusted to adapt to different emotional inputs and target styles without changing the original model structure.

[0061] LoRA technology inserts low-rank matrices into the Transformer layer of the pre-trained model for efficient fine-tuning. These low-rank matrices enable the model to quickly adapt to new tasks while maintaining the original structure, while reducing the computing resources and time required for training. This enables the art style model to efficiently generate artwork that matches the emotional input, with strong flexibility and adaptability.

[0062] S203, Emotion and style mapping optimization: Based on transfer learning technology, the generative model’s ability to generate diverse artistic styles is expanded, and a mapping mechanism from emotional parameters to diverse styles is constructed.

[0063] Specifically: Model training is carried out based on the existing Chinese heavy-color traditional Chinese painting database, and the model's ability to adapt to artistic styles is expanded through transfer learning. During the transfer learning process, the original traditional Chinese painting data is replaced with works of various artistic styles including oil painting and watercolor painting. The same process as traditional Chinese painting model training is adopted, including the use of Stable Diffusion technology, combined with low-rank adaptation and LoCon algorithm, to extract artistic style features.

[0064] S204, Automated Generation and Quality Optimization: Use the generative model to automatically generate artworks that meet the emotional parameters, and improve the quality and consistency of the generated artworks through optimization algorithms.

[0065] When generating artworks, an automated optimization algorithm is used to evaluate and optimize the quality of the generated artworks. The generated works will be automatically scored according to predetermined evaluation criteria (such as artistry, emotional expression, style consistency, etc.), thereby ensuring that the generated works have high quality. In order to further optimize the generation effect, this embodiment introduces a real-time feedback mechanism, and adjusts the emotional parameters according to the user's feedback on the generated works, so as to achieve the optimization and adjustment of the generated artworks. In the entire working process, the collaborative work of emotion recognition and work generation is the key. Through the real-time feedback mechanism, the user's emotional changes can timely affect the art generation process, so that the generated works are more in line with the user's emotional needs. For example, when the user's emotional state fluctuates violently, the generation strategy can be automatically adjusted to ensure that the style and emotional expression of the work are always consistent.

[0066] The artworks generated by the present invention include digital paintings, virtual reality art and other forms, which are adapted to the diverse emotional expression needs of users. The present invention can realize the full process automation of emotion recognition and artwork generation, solve the problems of insufficient emotion recognition accuracy and single art generation style in the prior art, and significantly improve the personalization of user emotional expression and the adaptability of art generation.

[0067] Example 2

[0068] This embodiment provides a specific example of the method for customized generation of artworks based on EEG emotion recognition described in Embodiment 1.

[0069] 1. Collect EEG data

[0070] The user's EEG activity is monitored in real time through high-precision EEG acquisition equipment to obtain the user's neural signals in specific emotional scenes. After processing, the EEG data is used to extract key features, such as the power of frequency bands related to emotional states (such as alpha waves, beta waves, etc.). These data provide a basis for the quantitative analysis of emotional states.

[0071] In the actual example, the following two steps are included:

[0072] 11: Collection of user emotional characteristics

[0073] In order to obtain the emotional characteristics of users, a rigorous experimental process is designed in this embodiment. First, 30 participants were recruited and divided into two groups according to age, with 20 people in the 20-30 age group and 10 people in the 31-50 age group. In order to ensure the diversity and representativeness of the data, 15 of the participants have professional art training backgrounds, and the other 15 have no relevant backgrounds. In terms of equipment configuration, the experiment used two EEG acquisition devices, Emotiv EPOC+ (14 electrodes, sampling frequency 128Hz) and NeuroSky (single electrode, sampling frequency 512Hz), which are used to capture multi-channel and single-channel brain wave signals respectively. At the same time, to ensure the quality of the experimental data, the experimental environment has been shielded from electromagnetic interference, the room noise is controlled below 40 decibels, and the light is stabilized between 300-500Lux.

[0074] In terms of experimental materials, 100 pictures were prepared as visual stimuli, covering three types: landscape, figure, and abstract art, with 30 pictures in each type. Among them, the landscape category is further divided into natural landscape and urban landscape, the figure category is divided into pictures that convey positive emotions and negative emotions, and the abstract art category conveys different emotional responses through color and composition. These designs are intended to trigger participants' diverse emotional responses and lay the foundation for subsequent emotional analysis and modeling.

[0075] 12: Quantification and labeling of emotional states

[0076] In the process of emotional scoring, this study used the standardized SAM scale to mark the user's emotional state. After viewing the pictures, each user scored the two dimensions of Arousal and Valence through a tablet device, with the score ranging from -1 to 1, accurate to two decimal places, to ensure the delicacy and accuracy of the results. Through this method, the user's emotional response was effectively quantified, providing reliable data support for the construction and verification of the subsequent model.

[0077] 2. Emotion Recognition

[0078] The sentiment analysis model processes the collected EEG feature data and extracts quantitative parameters such as the user's emotional intensity (Arousal value) and pleasure (Valence value) through a trained sentiment analysis algorithm. The sentiment analysis results accurately map the user's emotional state and provide customized emotional input for the subsequent art generation process.

[0079] It includes the following three steps:

[0080] 21. Preprocessing of sentiment data

[0081] In the data processing process, the collected EEG signals are first denoised to improve the quality of the data and the accuracy of subsequent analysis. Specifically, the independent component analysis (ICA) method is used to effectively remove artifact signals, such as interference caused by blinking or muscle movement. In addition, high-pass filtering (cutoff frequency is 0.5Hz) and low-pass filtering (cutoff frequency is 45Hz) are used to remove low-frequency drift and high-frequency noise, respectively, so as to retain the frequency components in the signal that are of analytical significance. After the denoising process is completed, the data is feature extracted, including three categories: frequency domain, time domain and spatial features. In the frequency domain features, the power spectral density of Alpha wave (8-12Hz), Beta wave (13-30Hz), Theta wave (4-7Hz) and Delta wave (0.5-3Hz) is extracted to reflect the signal strength in different frequency bands. In the time domain features, the mean, standard deviation and peak value of the signal are calculated to further capture the statistical characteristics of the signal. In the spatial features, the relative fluctuations between electrodes are extracted, such as the difference between the left frontal lobe and the right frontal lobe signals, to reveal the changes in the spatial distribution of the signals. The above steps provide rich feature data for the construction of the subsequent emotion model.

[0082] 22. Construction and training of emotion recognition model

[0083] In the construction and training of the emotion recognition model, the Random Forest Regressor was used as the main modeling method. The parameters of the model were set to 100 decision trees, with a maximum depth of 10, and the mean square error (MSE) was used as the splitting criterion. The data set was divided into a training set and a test set at a ratio of 80% and 20%, and the model performance was evaluated through 5-fold cross validation. The results showed that the average MSE of the model reached 0.03, R 2 The value is 0.92, which shows a high prediction accuracy and fitting effect. In order to further improve the robustness of the model, the model introduces L2 regularization to address the possible overfitting problem, thereby enhancing its generalization ability on different data samples.

[0084] 23. Output of emotional state

[0085] In the output stage of the emotional state, two floating-point numbers are generated in real time as emotional parameters, namely Arousal and Valence values. For example, the output format may be [Arousal: 0.85, Valence: 0.72], which represents the user's current emotional state. The Arousal value reflects the degree of emotional arousal, while the Valence value represents the degree of emotional pleasure. In addition, in order to enhance the credibility of the output, a confidence interval of the prediction result is also provided, such as [±0.05], indicating that the emotional state predicted by the model has a 95% confidence level within this interval. This output method provides quantitative and reliable emotional input for the subsequent artwork generation module, ensuring the personalization and accuracy of artistic creation.

[0086] 3. Artwork Generation

[0087] The emotion parameters (such as Arousal value and Valence value) output by the emotion recognition model are used to extract the target art style features (such as color, brushstrokes and composition). The model is fine-tuned by combining the pre-trained generative model and low-rank adaptation technology (LoRA) to generate digital artworks that match the user's emotions.

[0088] It includes the following three steps:

[0089] 31. Emotion Input and Style Extraction

[0090] In the emotional input and style extraction stage, a set of mapping rules are formulated based on emotional parameters (Arousal and Valence values) to generate an artistic style that matches the user's emotional state. Specifically, when both Arousal and Valence are high, bright colors (such as red and yellow) and strong color contrasts tend to be selected, and curves and dynamic compositions are often used in the works, showing a highly active and pleasant emotional state. If Arousal is high and Valence is low, cold colors (such as blue and gray) will be selected, the tones are dark, and the composition tends to be messy, reflecting the intense but negative tendency of emotions. For the emotional state of low Arousal and high Valence, soft tones (such as pink and green) will be used, and the composition will remain symmetrical and the elements will be sparse to express tranquility and pleasant emotions. Finally, if both Arousal and Valence are low, dark backgrounds and low-saturation colors will be used, with simple straight-line compositions, to convey a low and negative emotional experience. Through these mapping rules, the artwork can accurately reflect the user's emotional state and ensure the accuracy and personalization of customized creation.

[0091] 32. Generate model loading and fine-tuning

[0092] During the generation model loading and fine-tuning phase, a pre-trained art generation model was used. The pre-training dataset of the model contains 200,000 artworks, covering a variety of artistic styles such as impressionism, abstraction, and modern styles. In order to enable the model to generate customized artworks based on the user's emotional input, LoRA (low-rank adaptation) technology was used for fine-tuning. LoRA technology achieves rapid adaptation of the model by adjusting specific directions in the latent variable space and inserting low-rank matrices. During the fine-tuning process, the step size was set to 0.01, and convergence was achieved after 200 iterations. This fine-tuning process ensures that the model can generate artworks that meet user needs based on emotional input in a short period of time.

[0093] 33. Emotion and style mapping optimization

[0094] In the process of transfer learning, the model is first trained based on the existing Chinese heavy-color traditional Chinese painting database. The specific process is the same as the way of training the traditional Chinese painting model, that is, using the Stable Diffusion model and combining the low-rank adaptation (LoRA) technology and the LoCon algorithm for optimization. In order to expand the model's ability to recognize artistic styles, the data set was adjusted to replace the original traditional Chinese paintings with paintings of other artistic styles, such as oil paintings, watercolors, etc., so that the model can learn and generate artworks of different styles. In this way, a generative model that meets the requirements of a specific artistic style can be quickly trained. As a preliminary attempt, a data set containing more than 100 pictures of Chinese landscape paintings has been constructed, and a LoRA model specifically for generating landscape paintings has been successfully trained. This model is now able to generate artworks that conform to the style of Chinese landscape paintings.

[0095] 4. Robotic arm draws images

[0096] In the embodiment, the digitally generated artwork is physically drawn by a robotic arm. The robotic arm accurately controls the drawing process according to the color, line and texture information of the generated image, transforms the digital artwork into a real artwork, and satisfies the user's demand for the physicalization of customized artwork.

[0097] In terms of quality assessment, this embodiment uses predefined scoring rules to evaluate the generated artworks from three dimensions: artistry (0-10 points), consistency of emotional expression (0-5 points), and degree of style matching (0-5 points). If the average score of the generated artwork exceeds 13.5 points, it is considered a qualified artwork. In addition, users can subjectively score the generated artworks, and adjust the emotion and style mapping parameters in real time based on user feedback to achieve continuous optimization.

[0098] 5. Experimental results

[0099] 51. Emotion Recognition

[0100] During the experiment, participants were invited to watch a colorful and vibrant picture of a seaside sunrise to accurately induce their emotional responses. High-precision EEG acquisition equipment was used in the experiment to obtain the participants' original EEG signals. By analyzing the frequency band power spectral density of the signal, it was found that the α wave band power was relatively low, the β wave band power was high, and the γ wave band was also active. Combined with the emotion recognition model, the emotional state of the participants was predicted, and the output Valence value was 0.95 and the Arousal value was 0.98, indicating that they were in a highly pleasant and highly aroused emotional state.

[0101] 52. Specific style image generation

[0102] According to the results of emotion prediction, Chinese heavy-color traditional Chinese paintings were selected as the core painting category for image generation. Since the Valence and Arousal values ​​were close to 1, the GLM-4 large language model generated the Chinese heavy-color traditional Chinese painting style feature keywords that matched the emotions: dynamic, energetic, bold, vivid, intense, passionate, striking, powerful, fiery, expressive. These keywords were input into the trained LoRA model. The model combines the improved generative adversarial network (GAN) algorithm, the low-rank adaptation (LoRA) technology and the LoCon algorithm, and learns and extracts the macro style and micro detail features of traditional Chinese paintings through the combination of low-rank matrix decomposition and convolution operations. Finally, an image with a strong Chinese heavy-color traditional Chinese painting style was generated. The main colors of the picture are bright, mainly red, orange, yellow and blue, supplemented by delicate small lines, which vividly convey the emotional state of the participants when viewing the picture, and the effect is consistent with expectations.

[0103] This embodiment verifies that the present invention can generate high-quality artworks based on user emotions and is suitable for various application scenarios such as customized art creation.

[0104] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning, or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for customized generation of artworks based on EEG emotion recognition, characterized in that: The following steps are involved: S1, by collecting the user's neural signals, extracting feature data related to the emotional state, and converting the feature data into emotional parameters representing the intensity and pleasure of emotion through a quantitative model; S2, dynamically matches the target artistic style characteristics based on the emotional parameters, and uses the generative model to automatically generate artworks.

2. The method for customizing artworks based on EEG emotion recognition according to claim 1, characterized in that: The S1 comprises the following steps: S101, collecting user emotional features: obtaining the user's immediate emotional state through neural signals to form a multi-dimensional feature matrix; S102, quantification and labeling of emotional state: based on the constructed emotional evaluation system, quantify the user's emotional experience, obtain emotional parameters, label the multi-dimensional feature matrix, and form an emotional feature data set; S103, preprocessing of emotional data: cleaning, denoising and normalizing the emotional feature data set; S104, construction and training of emotion recognition model: construct an emotion recognition model using machine learning algorithm, train the emotion recognition model based on the emotion feature dataset, and learn the mapping relationship between neural signals and emotion parameters; S105, output of emotional state: based on the emotion recognition model, generate real-time quantitative parameters of the user's emotional state.

3. The method for customized generation of artworks based on EEG emotion recognition according to claim 1, characterized in that: The neural signal includes an electroencephalogram signal, specifically, a power spectrum density value in different frequency bands.

4. The method for customizing artwork based on EEG emotion recognition according to claim 2, characterized in that: The quantification of the user's emotional experience is specifically as follows: converting the user's emotional state into an Arousal value and a Valence value, wherein the Arousal value represents the degree of arousal and the Valence value represents the degree of pleasure.

5. The method for customizing artworks based on EEG emotion recognition according to claim 4, characterized in that: The S2 comprises the following steps: S201, emotion input and style extraction: dynamically select the target art style features based on the emotion parameters and extract the core elements of the artwork generation; S202, generative model loading and fine-tuning: quickly fine-tune the pre-trained generative model through low-rank adaptation technology to adapt it to the target artistic style; S203, Emotion and style mapping optimization: Based on the transfer learning technology, the generation model's ability to generate diverse artistic styles is expanded, and a mapping mechanism from emotion parameters to diverse styles is constructed; S204, Automated Generation and Quality Optimization: Use the generative model to automatically generate artworks that meet the emotional parameters, and improve the quality and consistency of the generated artworks through optimization algorithms.

6. The method for customizing artworks based on EEG emotion recognition according to claim 5, characterized in that: The dynamic selection of the target art style features includes extracting key color features, brushstroke features and composition layout features in the artwork, and the style features are adaptively selected according to the user's emotional intensity and emotional tendency, wherein the Arousal value and Valence value are divided into two levels, high and low, according to a preset threshold. When both the Arousal value and the Valence value are high, choose bright colors and strong color contrast, use more curves and dynamic composition, and express a highly active and pleasant emotional state; when the Arousal value is high and the Valence value is low, choose cool colors, dark tones, and a messy composition, reflecting an intense but negative emotional tendency; when the Arousal value is low and the Valence value is high, use soft tones, keep the composition symmetrical and the elements sparse, to express tranquility and pleasant emotions; when both the Arousal value and the Valence value are low, use a dark background and low-saturation colors, with a simple straight line composition, to convey a low and negative emotional experience.

7. The method for customizing artwork based on EEG emotion recognition according to claim 5, characterized in that: In the generative model loading and fine-tuning, the pre-trained GLM-4 large language model is used, and the low-rank adaptation technology is used to insert a low-rank matrix into the Transformer layer of the pre-trained model to perform model fine-tuning.

8. The method for customizing artworks based on EEG emotion recognition according to claim 5, characterized in that: Specifically, S203 is as follows: model training is performed based on the existing Chinese heavy-color traditional Chinese painting database, and the model's ability to adapt to artistic styles is expanded through transfer learning. During the transfer learning process, the original traditional Chinese painting data is replaced with works of various artistic styles including oil painting and watercolor painting. The same process as the traditional Chinese painting model training is adopted, including the use of Stable Diffusion technology, combined with low-rank adaptation and LoCon algorithm, to extract artistic style features.

9. The method for customizing artworks based on EEG emotion recognition according to claim 5, characterized in that: In S204, a real-time feedback mechanism is introduced to adjust the emotional parameters according to the user's feedback on the generated work, thereby achieving optimization and adjustment of the generated art work.

10. The method for customizing artwork based on EEG emotion recognition according to claim 1, characterized in that: The artworks include digital paintings and virtual reality art.

Images