A multi-dimensional semantic association-based image fragmentation reorganization method

Abstract

The application discloses a multi-dimensional semantic association image fragmentation reorganization method, comprising the following steps: S1, receiving a user uploaded image file and input text theme, and analyzing a core theme; S2, for each input image, using a deep learning model to analyze content, style, emotion and time, and labeling a set of multi-dimensional semantic labels; S3, according to the analyzed core theme, grouping and sorting the labeled image set according to a logical relationship; S4, providing a fragmentation template, performing fragmentation reorganization based on semantic matching, and mapping the sorted images to each fragmentation area of the template; S5, according to the matching result, performing image cropping, transformation and synthesis to generate a final composite image. The application can solve the technical problems in the prior art that a single image or simple collage is difficult to carry multiple narratives and complex connotations, and the creation process is heavily dependent on manual work and lacks intelligent semantic association.

Description

Technical Field

[0001] This invention relates to an image segmentation method, and more particularly to an image fragmentation and reconstruction method with multidimensional semantic association. Background Technology

[0002] In today's digital media age, images are a core medium for information transmission and emotional expression. However, a single still image has inherent limitations in carrying complex narratives, multiple perspectives, or time spans. To overcome this limitation, existing technologies mainly employ the following methods for image combination and processing:

[0003] Manual image collage and compositing: This is the most traditional and common method. Artists or designers use professional image editing software, such as Adobe Photoshop, to manually combine elements of multiple images using tools such as layers, masks, and blending modes. This method can create highly artistic and expressive works, but it heavily relies on the operator's professional skills, artistic literacy, and a significant time investment. The creative process is complex and difficult to automate and scale.

[0004] Automated Collage Tools: Numerous mobile applications and online tools (such as Canva and PicCollage) on the market can quickly arrange multiple images into a preset grid template. These tools are simple to use and efficient, but their core logic is based on geometric layout and space filling, rather than the inherent meaning of the image content. The generated collages are usually just simple listings of images, lacking deep semantic connections and narrative logic, and are difficult to express complex themes.

[0005] Photo Mosaic: This technique stitches together hundreds or thousands of small images to form a large target image. Its focus is on using the color and brightness information of the smaller images to simulate the pixels of the large image, thus creating a visual effect on a macroscopic scale. While technically achieving multi-image combination, its focus is on pixel-level color matching; the semantic relationships between the smaller images and their narrative connection to the macroscopic theme are often ignored.

[0006] Image style transfer: A deep learning-based technique, as described in Gatys et al.'s paper "A Neural Algorithm of Artistic Style" (reference: Gatys, LA, Ecker, AS, & Bethge, M. (2015). A Neural Algorithm of Artistic Style. arXiv preprint arXiv:1508.06576), applies the artistic style of one image to the content of another. This is a separation and recombination of content and style, but it still operates between two images and does not achieve complex recombination of multiple images across multiple semantic dimensions to construct a new narrative structure.

[0007] As can be seen from the above, the existing technology has the following defects or shortcomings:

[0008] Lack of semantic depth and narrative ability: Existing automated tools primarily address the visual problem of "how to arrange" images, neglecting the semantic problem of "why arrange them this way." They cannot understand the story, emotions, or logical relationships behind the image collection, and therefore cannot generate composite images with deep narrative capabilities.

[0009] The creative process is highly dependent on human intervention: to create complex, multi-layered images, the meticulous manual operation of professional designers is still indispensable. The entire process is time-consuming and labor-intensive, with a high barrier to entry, and lacks an effective method that can understand creative intent and assist in intelligent creation.

[0010] Limited Dimensionality of Association: Existing image composition techniques typically only consider the association of visual elements (color, layout) or single content (such as "all images containing faces"). They cannot simultaneously process and associate information across multiple dimensions, such as style, time, emotion, and content elements, and therefore cannot achieve complex visual expressions such as "multi-dimensional deconstruction of the same thing" or "emotional contrast across time and space".

[0011] Limited expressive paradigms: Whether it is a simple grid puzzle or a photo mosaic, the final presentation is a relatively flat aggregation of information, making it difficult to achieve a deep visual narrative with "multiple truths" or multiple perspectives in a single work. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to provide a multi-dimensional semantic association image fragmentation and reconstruction method, which can solve the technical problems in existing image processing technology, such as the inability of a single image or simple collage to carry multiple narratives and complex connotations, and the heavy reliance on manual labor and lack of intelligent semantic association in the creation process.

[0013] To address the aforementioned technical problems, this invention provides a multi-dimensional semantic association method for image fragmentation and reconstruction, comprising the following steps: S1, receiving image files uploaded by a user and input text topics, and parsing the core topics; S2, for each input image, performing content, style, sentiment, and time analysis using a deep learning model, and labeling it with a set of multi-dimensional semantic tags; S3, grouping and sorting the labeled image set according to logical relationships based on the parsed core topics; S4, providing fragmentation templates, performing fragmentation and reconstruction based on semantic matching, and mapping the sorted images to various fragment regions of the template; S5, cropping, transforming, and synthesizing the images according to the matching results to generate the final composite image.

[0014] Furthermore, S1 performs natural language processing on the core topic text to extract keywords, topic concepts, and potential narrative structures, and introduces the following context-aware semantic weight matrix:

[0015] ; t represents the target word, c represents the cultural / regional context, s represents the complete sentence, α, β, and γ are weight parameters learned from training data, BERT(t) represents the word vectors of the pre-trained model, LocalDict(t,c) is a localized dictionary used to map the target word t to a more accurate base word based on the context c, and ContextEmb(t,s) is a context embedding vector generated by analyzing the role of word t in the entire sentence s.

[0016] Furthermore, S2 represents each image as a semantic vector containing multiple key-value pairs, including: content element annotation: using object detection models and image classification models to identify key objects, scenes, and activities in the image; style feature annotation: using convolutional neural networks to classify the artistic style of the image; sentiment annotation: using a model trained on a sentiment image dataset to analyze the overall emotional atmosphere conveyed by the image; and time attribute annotation: extracting the shooting time from the image's metadata, or inferring its relative time attribute through image content analysis.

[0017] Furthermore, S2 rapidly evaluates the basic image quality score from the dimensions of sharpness, noise, and contrast. Q base (I) : Q base (I)= w_s × S(I) + w_n × N(I) + w_c × C(I); where w_s, w_n, w_c are weighting coefficients, their sum is 1, S(I) is the normalized sharpness score, N(I) is the normalized noise score, and C(I) is the normalized contrast score; when Q base (I) Images below a preset threshold are considered low-quality or blurry. ， Further fine calibration is performed using the deep learning quality evaluation function Q(I); when processing low-quality or blurry images, S2 employs the following adaptive confidence calibration algorithm:

[0018] ;

[0019] Where: Q(I) is the image quality evaluation function: T(age) is the time decay correction factor: , where age represents the year of the image.

[0020] Furthermore, S3 employs the following logical relationships for grouping and sorting: Multidimensional deconstruction of the same thing: Images with the same core content element tags are grouped together and then sorted according to style, time, or emotion dimensions; the cosine similarity between the style feature vectors of all images within the same group is maximized to improve style uniformity; Continuous narrative: Images are linearly sorted according to time attributes or the causal logic of events to construct a story flow; the total variance of the time attributes of all images in the story flow is minimized to ensure smooth arrangement in the time dimension; Contradiction and contrast: Euclidean distance is used to calculate the emotional contrast of two images, and images with opposite or conflicting emotional / content tags are paired or grouped.

[0021] Furthermore, S3 also includes introducing the following term into the layout optimization function by adding narrative adjacency constraints to spatial layout conflicts, such that logically adjacent images are spatially adjacent:

[0022] ;

[0023] Where, Pos(i) and Pos(j) Representing images respectively i and images j The center coordinates of the fragment region assigned in the final composite image layout.

[0024] Further, S4 includes: each fragmentation template is divided into multiple independent fragment regions; each fragment region has position, shape and size attributes, and has at least one or more semantic anchors; for a certain fragment region, the image that best matches its semantics is found in the image set generated in step S3 according to its semantic anchors; after finding it, the content of the image is cropped and filled into the fragment region.

[0025] Furthermore, in step S4, for multiple candidate images appearing at the same semantic anchor point, the following multi-objective scoring function is designed, and the image with the highest score is automatically selected:

[0026] ;

[0027] in For candidate image features, w is the ideal vector required for the anchor point. style For style dimension, w emotion It is the emotional dimension, and the parameters α, β, and γ represent the weights that the user adjusts.

[0028] Furthermore, S4 also includes the following multi-level fallback strategy: Level 0, Optimal Path: When the confidence score of key content element detection is greater than 0.95 and the semantic similarity of the retrieved matching content is greater than 0.85, the complete style and sentiment transfer process is executed; Level 1, Downgraded Matching: When the similarity of the matched face is between 0.6 and 0.85, a Top-K weighted fusion strategy is adopted, and the generation intensity parameter is reduced; the generation intensity parameter includes the style transfer intensity coefficient, the sentiment rendering weight, and the feature fusion ratio; the value range of the style transfer intensity coefficient is 0.3 to 0.7, the value range of the sentiment rendering weight is 0.2 to 0.6, and the value range of the feature fusion ratio is 0.4 to 0.8; Level 2, Generalized Template: When the retrieval fails but the key content element detection is successful, a predefined style / sentiment template library is used; Level 3, Minimal Intervention: When the generation module fails, it falls back to the traditional image processing method; Level 4, Original Output: When the input image cannot be processed, the original image is returned with additional diagnostic information.

[0029] Further, S5 includes: S51, optimal window cropping: First, the most informative region in the image is located using a saliency detection model. Then, based on the aspect ratio of the fragments, a window that maximizes the inclusion of salient content is selected for cropping; S52, local perceptual geometric transformation: When mapping the cropped rectangular window to the non-rectangular fragment region, a deformation penalty weight is applied to the key feature points in the image using a local perceptual constraint transformation algorithm. This prioritizes maintaining the proportion and shape of these key regions and allows secondary background regions to generate more deformation. The deformation penalty weight λ_deform ranges from [0.1, 1.0], or is determined based on the basic image quality score. Q base (I) Dynamic adjustment is performed, and the calculation formula is: λ_deform(I) = 0.2 + 0.6 · (1 - Q base (I) S53, Boundary Processing and Generative Fill: After all fragments are filled, the seams between the fragments are processed; when the cropped image cannot completely cover the fragment area, the generative fill module is enabled, and new pixels are generated based on the content of the surrounding filled pixels using a diffusion model or generative adversarial network to seamlessly fill the blank area; S54, Global Style Coordination: Finally, a global color and lighting analysis is performed on all filled fragments, and fragments whose hue and contrast deviate from the overall atmosphere by more than a preset threshold are automatically corrected.

[0030] Compared with the prior art, the present invention has the following beneficial effects: The multi-dimensional semantic association image fragmentation and reconstruction method provided by the present invention provides a method that can automatically analyze and understand the semantics of multiple images in multiple dimensions such as content, style, emotion, and time.

[0031] Establish a logical relationship-based intelligent image grouping and sorting mechanism to enable image combinations to go beyond simple visual arrangement and serve specific narrative goals.

[0032] By combining semantic association with visual fragmentation, a novel composite image form is created, which is visually presented as fragments, but highly unified and related semantically, thereby achieving a deep visual narrative.

[0033] In practical implementation, this invention introduces multi-dimensional semantic vector space modeling. Each image is embedded into a high-dimensional feature space through multimodal representation learning, where the semantic dimension corresponds to the coordinate axes of a mathematical vector. The logical relationships of image combinations are formalized into mathematical constraints or optimization objectives. This includes:

[0034] Temporal continuity optimization: Minimize the total variance of the combined image on the "time axis";

[0035] Enhanced emotional contrast: Maximizes the Euclidean distance projected between segments on the "emotional dimension";

[0036] Style uniformity constraint: For the same logical group, minimize the cosine similarity difference between its style feature vectors. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the implementation of image fragmentation and reconstruction based on multidimensional semantic association in this invention.

[0038] Figure 2 This is a diagram of the image fragmentation and reconstruction system architecture for multidimensional semantic association according to the present invention. Detailed Implementation

[0039] The present invention will now be further described with reference to the accompanying drawings and embodiments.

[0040] Figure 1 This is a flowchart illustrating the implementation of image fragmentation and reconstruction based on multidimensional semantic association in this invention.

[0041] Please see Figure 1 The present invention mainly includes the following steps:

[0042] Step 1: Data reception and topic parsing;

[0043] The system receives a core topic input from the user (e.g., a descriptive text such as "a century of change in a city") and a set (multiple) of input images. The system performs Natural Language Processing (NLP) on the core topic text to extract keywords, topic concepts, and potential narrative structures (such as timelines, contrast relationships, etc.).

[0044] In practical implementation, this invention discovered a key technical challenge: semantic ambiguity in multilingual environments. When user input contains technical terms, dialects, or descriptions with cultural background, traditional NLP models often exhibit comprehension biases.

[0045] For example, when a user enters "the nostalgia of old Shanghai's alleyways", the regionally specific word "alleyway" may be incorrectly identified as "hutong" or "alley" in a standard dictionary, leading to deviations in subsequent semantic associations.

[0046] This invention introduces a context-aware semantic weight matrix:

[0047] ;

[0048] t represents the target vocabulary; c represents the cultural / regional context; s represents the complete sentence; α, β, and γ are weight parameters, learned from training data; BERT(t) represents the word vectors of the pre-trained model.

[0049] LocalDict(t,c) is a localization dictionary used to map the target word t to a more precise base word in the context c;

[0050] ContextEmb(t,s) is a context embedding vector generated by analyzing the role of word t in the whole sentence s;

[0051] Using this method, the system's accuracy in understanding "lane" increased from 62.3% to 89.7%.

[0052] Step 2: Multidimensional semantic annotation;

[0053] For each input image, a deep learning model is used for analysis, and a set of multi-dimensional semantic labels are assigned to it. Ultimately, each image is represented as a multi-dimensional semantic vector V_img = {V_content, V_style, V_emotion, V_time, ...}. This process specifically includes:

[0054] Content element annotation: Use object detection models (such as YOLO - You Only Look Once or Faster R-CNN) and image classification models (such as ResNet) to identify key objects, scenes and activities in the image (e.g., "buildings", "streets", "festivals").

[0055] Style feature annotation: Use a specific convolutional neural network (CNN) to classify the artistic style of the images (such as "black and white", "retro", "cyberpunk", "watercolor").

[0056] Sentiment annotation: Using a model trained on a sentiment image dataset (such as AffectNet), analyze the overall emotional atmosphere (such as "joy", "sadness", "nostalgia", "tranquility") conveyed by the image.

[0057] Time attribute annotation: The shooting time is extracted first from the image's metadata (such as EXIF ​​information). If there is no metadata, the relative time attribute (such as "past", "modern", "futuristic") is inferred through image content analysis (such as clothing, architectural style, image aging degree).

[0058] Output: Each image is represented as a semantic vector containing multiple key-value pairs, for example: Image A: {Content:'Old Building', Time: 'Past', Emotion: 'Nostalgia', Style: 'Black and White'}.

[0059] When processing low-quality or blurry images, the accuracy of traditional object detection models drops sharply. Especially for old historical photographs, due to issues such as image noise, low contrast, and limited resolution, the confidence level of object detection often falls below the threshold.

[0060] This invention uses a multi-index weighted quality assessment function to determine image quality. Q base (I) Defined as a weighted sum of several basic image quality metrics, which are evaluated from the dimensions of sharpness, noise, and contrast respectively:

[0061] Q base (I) = w_s × S(I) + w_n × N(I) + w_c × C(I);

[0062] Where w_s, w_n, w_c are weighting coefficients, and their sum is 1 (for example, they can be set to 0.5, 0.3, 0.2), and S(I), N(I), C(I) are the normalized sharpness, noise, and contrast scores, respectively.

[0063] i. Sharpness / Blurriness Index S(I): Highly sensitive to edges and details (high-frequency components) in an image, this index effectively quantifies the degree of blurriness. It is simple and efficient to calculate, making it ideal as a core metric for real-time evaluation. For a sharp image, rich edge details result in a large range of pixel value variations (i.e., variance) after Laplacian processing. Conversely, a blurred image has a small variance.

[0064] As a preferred embodiment, the index is calculated as follows:

[0065] 1. Convert the input image I to a grayscale image I_gray.

[0066] 2. Perform a convolution operation on I_gray using a 3×3 Laplacian kernel (as shown below) to obtain the response map L_map:

[0067] Kernel = [[0,1,0],

[0068] [1, -4,1],

[0069] [0,1,0]].

[0070] 3. Calculate the variance Var(L_map) of all pixel values ​​in the response map L_map.

[0071] 4. S(I) is the variance value, which is normalized to the [0, 1] interval using a preset mapping function (such as the Sigmoid function). The larger the variance, the closer S(I) is to 1, indicating a clearer image.

[0072] ii. Noise Level Index N(I): Based on local variance assessment, this index can distinguish normal texture details from meaningless noise, providing a reliable quantitative basis for noise levels. Noise in images typically manifests as random, high-frequency fluctuations in pixel values. In smooth areas of an image (such as the sky or walls), ideally, pixel values ​​should be very close, with small variance. If the local variance in these areas is large, it indicates the presence of noise.

[0073] As a preferred embodiment, the index is calculated as follows:

[0074] 1. First, use an edge detection algorithm (such as Canny or Sobel) to identify the non-edge regions (smooth regions) of the image.

[0075] 2. Within these smooth areas, randomly select multiple small image patches (e.g., 8×8 pixels).

[0076] 3. Calculate the variance of pixel values ​​for each image block and then calculate its average value, AvgVar_smooth.

[0077] 4. N(I) is inversely proportional to AvgVar_smooth. It is normalized using N(I) = 1 - min(AvgVar_smooth / T_noise, 1), where T_noise is a noise tolerance threshold. The lower the noise, the closer N(I) is to 1.

[0078] iii. Contrast Metric C(I): Root Mean Square Contrast (RMS Contrast). Contrast reflects the degree of difference between bright and dark areas in an image. This metric effectively assesses the overall dynamic range and sharpness of an image, avoiding detail loss due to underexposure or over-aging. Low-contrast images (hazy) are generally considered low-quality. RMS contrast measures global contrast by calculating the standard deviation of pixel intensity relative to the average intensity.

[0079] As a preferred embodiment, the index is calculated as follows:

[0080] 1. Convert the input image I to a grayscale image I_gray.

[0081] 2. Calculate the average intensity μ_gray of all pixels.

[0082] 3. Calculate RMS contrast: RMS = sqrt( (1 / MN) × Σ( (p_ij - μ_gray)^2 ) ), where M and N are the image size and p_ij is the pixel intensity.

[0083] 4. The RMS value is non-linearly mapped through a preset ideal range (e.g., the RMS value of health contrast is between 40 and 60) to obtain the normalized C(I) score.

[0084] This invention develops an adaptive confidence calibration algorithm:

[0085] ;

[0086] Where: Q(I) is the image quality evaluation function based on a deep learning model: BRISQUE (Blind / Referenceless Image Spatial Quality Evaluator) is a blind / referenceless image spatial quality evaluator; NIQE (Natural Image Quality Evaluator) is a natural image quality evaluator; T(age) is the time decay correction factor. , where age is the year of the image.

[0087] Image quality baseline score Q base (I) The advantages of this approach are high computational efficiency, suitability for real-time processing, and flexible adjustment of weights across dimensions based on application requirements. The disadvantages are that it only evaluates three predefined dimensions and has limited coverage of complex distortion types. The advantages of Q(I) are that it is based on statistical learning of natural scenes, can more comprehensively evaluate image quality, and has a higher correlation with human subjective perception. The disadvantages are higher computational complexity and unsuitability for real-time scenarios requiring low latency. In practical deployments, the image quality evaluation function Q(I) is used for offline processing with high quality requirements; for online processing with high real-time requirements, the image quality evaluation function Q(I) is used first. Q base (I)， when Q base (I) Images below a preset threshold are identified as low-quality or blurry, and then further fine-tuned using the deep learning quality evaluation function Q(I).

[0088] Function selection rules: In each step of the process, Q base (I)Q(I) is used for initial screening and real-time judgment (such as the initial quality screening in step S2 and the dynamic parameter adjustment in step S5), and is used for scenarios that require precise calibration (such as the confidence calibration for images that have been judged to be of low quality in step S2).

[0089] In the semantic annotation step, more fine-grained tags can be introduced, such as image composition (symmetry, rule of thirds), lighting (highlights, shadows), and color (dominant hue, complementary colors), to achieve more precise control.

[0090] Step 3: Intelligent grouping and sorting driven by logical relationships;

[0091] This invention introduces a quantized evaluation function in the image grouping, sorting, and matching process. Assume a composite image consists of N image fragments {I1, I2, ..., I...} n Composed of, each fragment I i Corresponding to a multidimensional semantic vector V i = {v i _content, v i _style, v i _emotion, v i _time}.

[0092] Based on the core themes or user-preset / custom logical relationships parsed in Step 1, the labeled image set is intelligently grouped and sorted. Logical relationships serve as the "script" guiding image reconstruction; common logical relationships include:

[0093] "Multidimensional Deconstruction of the Same Thing": Images with the same core "content element" tag are grouped together and then sorted according to other dimensions such as "style," "time," or "emotion." For example, for the theme "A Person's Life," the system will group all images containing that person and sort them by the "time" tag.

[0094] When presenting the same topic from different angles, it may be necessary to maintain stylistic consistency to enhance the overall coherence. This can be achieved by maximizing the cosine similarity between the style feature vectors v_style of image fragments within the same logical group. For a logical group G containing M images, its style consistency score can be defined as the average cosine similarity of all image pairs within the group:

[0095] ;

[0096] When grouping and selecting images, the system tends to choose the combination of images that maximizes the score.

[0097] “Continuous Narrative”: This involves linearly arranging images based on the “time” attribute or the causal logic of events to construct a story flow. For example, “A Century of Change in a City” aims to smoothly arrange image fragments along the time dimension. This can be achieved by minimizing its time attribute t. i It is achieved through the total variance.

[0098] Step 3.1: Calculation of the total variance of time continuity:

[0099] When employing a "continuous narrative" logical relationship, the system needs to quantify the smoothness of the image sequence over time. The specific implementation steps are as follows:

[0100] Step 3.1.1: Quantification of time attributes;

[0101] For images containing EXIF ​​timestamps, directly extract the year information, for example: t1= 1985, t2= 1990, t3= 2020;

[0102] For images without metadata, the relative time value is output by the time inference model and mapped to the standardized interval [0,100].

[0103] Step 3.1.2: Calculate the total variance;

[0104] The formula for calculating the total variance is:

[0105] Var(T) = (1 / N) × Σ i=1 (t i - t)²

[0106] Where: N is the total number of images; t i Let t be the temporal attribute value of the i-th image; t = (1 / N) × Σ i=1 t i This is the average of all time values.

[0107] Implementation example:

[0108] Suppose there are 5 images with time attributes as follows: T = {1950, 1970, 1990, 2010, 2020};

[0109] First step, calculate the average:

[0110] t= (1950 + 1970 + 1990 + 2010 + 2020) / 5 = 1988;

[0111] The second step is to calculate the squared deviation of each time point from the average:

[0112] (1950-1988)² = 1444;

[0113] (1970-1988)² = 324;

[0114] (1990-1988)² = 4;

[0115] (2010-1988)² = 484;

[0116] (2020-1988)² = 1024;

[0117] The third step is to calculate the total variance:

[0118] Var(T) = (1444 + 324 + 4 + 484 + 1024) / 5 = 656;

[0119] Step 3.1.3: Sorting optimization;

[0120] The system employs a greedy algorithm or dynamic programming method to select the arrangement that minimizes Var(T) from all possible image permutations. A smaller variance indicates a more concentrated or uniform temporal distribution, and stronger narrative continuity.

[0121] The optimization objective can be formalized as follows:

[0122] ;

[0123] Where T = {t1, t2, ..., t} n} is a set of time tags (which can be quantified, such as years) corresponding to image fragments. This is the average value of this set of time labels. During the sorting step, the system prioritizes image sequences that minimize this loss function.

[0124] Step 3.2: Calculate the Euclidean distance for emotional contrast;

[0125] When employing the logical relationship of "contradiction and contrast," the system needs to quantify the degree of emotional difference between images. The specific implementation steps are as follows:

[0126] Step 3.2.1: Vectorization of sentiment features;

[0127] Using a deep learning model trained on sentiment datasets such as AffectNet, the sentiment of each image is mapped to a D... e A dimensional vector space. Common sentiment dimensions include:

[0128] Valence: Range [-1, 1], negative values ​​indicate negativity, and positive values ​​indicate positivity;

[0129] Arousal: Range [0, 1], representing the intensity of emotion;

[0130] Dominance: Range [0, 1], representing the degree of control felt;

[0131] For example, image I i The emotion vector can be represented as:

[0132] v_emotion_i = [valence_i, arousal_i, dominance_i];

[0133] Specific examples:

[0134] Image A (Joyful Celebration): v_emotion_A = [0.8, 0.9, 0.7];

[0135] Image B (ruins scene): v_emotion_B = [-0.6, 0.3, 0.2].

[0136] Step 3.2.2: Calculate the Euclidean distance;

[0137] For two images I that need to be compared i and I j The emotional contrast is calculated using the Euclidean distance formula: ;

[0138] Where: D e The dimension of the sentiment vector (3 in this example); v ik For image I i The value at the k-th sentiment dimension; v jk For image I j The value at the k-th sentiment dimension;

[0139] Implementation example:

[0140] Calculate the emotional contrast between image A and image B above:

[0141] D_emotion(A, B) = √[(0.8-(-0.6))² + (0.9-0.3)² + (0.7-0.2)²];

[0142] = √[1.4² + 0.6² + 0.5²];

[0143] = √[1.96 + 0.36 + 0.25];

[0144] = √2.57;

[0145] ≈ 1.60;

[0146] Step 3.2.3: Compare the pairing strategies;

[0147] The system selects the image pair with the largest emotional Euclidean distance from all candidate images for pairing to maximize visual impact. In practical applications, a distance threshold D_threshold is set; only when D_emotion > D_threshold are the two images considered to have sufficient contrast.

[0148] For cases requiring pairing M groups of images for comparison, a maximum weighted bipartite graph matching algorithm (such as the Hungarian algorithm) is used to solve the problem: maximize Σ m=1 D_emotion(I am , I m );

[0149] "Contradiction and Contrast": Pair or group images with opposite or conflicting "emotions" (such as "joy" and "sadness") or "contents" (such as "nature" and "industry") to highlight the tension of the theme.

[0150] For example, in the case of "Industry and Nature," the goal is to maximize the differences in the emotional dimension among image fragments with conflicting labels. This can be achieved by maximizing the Euclidean distance between their emotional feature vectors v_emotion. For a pair of image fragments I to be compared... i and I j Its emotional contrast score can be defined as:

[0151] ;

[0152] Where v_emotion is a D e A 3D vector represents coordinates in the emotion space. During grouping, the system will prioritize pairing images with higher contrast scores.

[0153] Step 3.3: Calculate the cosine similarity of style consistency;

[0154] When employing the logical relationship of "multi-dimensional deconstruction of the same thing," the system needs to ensure stylistic consistency among images within the same group. The specific implementation steps are as follows:

[0155] Step 3.3.1: Style Feature Extraction;

[0156] Style features of images are extracted using a pre-trained convolutional neural network (such as VGG-19). The specific method is as follows:

[0157] Input the image into the network;

[0158] Extract feature maps from multiple convolutional layers;

[0159] Calculate the Gram matrix for each feature map layer;

[0160] Flatten and concatenate all Gram matrices to form style feature vectors.

[0161] For example, after extracting the features from three layers (conv1_1, conv2_1, conv3_1), image I... i The style vector can be represented as: v_style_i = [gram1, gram2, gram3] ∈ s ;

[0162] Where D s The total dimensions of style features (usually thousands of dimensions).

[0163] Step 3.3.2: Calculate cosine similarity;

[0164] For two images I i and I j The style similarity is calculated using the cosine similarity formula:

[0165] CosineSim(I i , I j ) = (v_style_i · v_style_j) / (||v_style_i|| × ||v_style_j||);

[0166] in:

[0167] v_style_i · v_style_j represents the dot product of two vectors: ;

[0168] ||v_style_i|| represents the L2 norm of vector v_style_i: √ ;

[0169] ||v_style_j|| represents the L2 norm of vector v_style_j: √ .

[0170] The cosine similarity value ranges from [-1, 1], where:

[0171] 1 indicates that the two vectors have the same direction (completely consistent style);

[0172] 0 indicates that the two vectors are orthogonal (style irrelevant);

[0173] -1 indicates that the two vectors have completely opposite directions (completely opposite styles).

[0174] Implementation example:

[0175] Assuming a simplified 3D style vector:

[0176] Image A (black and white photography): v_style_A = [0.8, 0.3, 0.1];

[0177] Image B (black and white photography): v_style_B = [0.7, 0.4, 0.15];

[0178] Image C (oil painting style): v_style_C = [0.2, 0.1, 0.9];

[0179] Calculate the cosine similarity between images A and B:

[0180] First step, calculate the dot product:

[0181] v_A · v_B = 0.8×0.7 + 0.3×0.4 + 0.1×0.15 = 0.56 + 0.12 + 0.015 =0.695;

[0182] The second step is to calculate the norm:

[0183] ||v_A|| = √(0.8² + 0.3² + 0.1²) = √(0.64 + 0.09 + 0.01) = √0.74 ≈0.860;

[0184] ||v_B|| = √(0.7² + 0.4² + 0.15²) = √(0.49 + 0.16 + 0.0225) = √0.6725 ≈ 0.820;

[0185] The third step is to calculate the cosine similarity:

[0186] CosineSim(A, B) = 0.695 / (0.860 × 0.820) ≈ 0.695 / 0.705 ≈ 0.986

[0187] Similarly, calculate the cosine similarity between images A and C:

[0188] v_A · v_C = 0.8×0.2 + 0.3×0.1 + 0.1×0.9 = 0.16 + 0.03 + 0.09 =0.28;

[0189] ||v_C|| = √(0.2² + 0.1² + 0.9²) = √(0.04 + 0.01 + 0.81) = √0.86 ≈0.927;

[0190] CosineSim(A, C) = 0.28 / (0.860 × 0.927) ≈ 0.28 / 0.797 ≈ 0.351;

[0191] The results show that the style similarity between images A and B is 0.986 (highly similar), while the style similarity between images A and C is only 0.351 (low similarity).

[0192] Step 3.3.3: Assessment of style consistency within the group;

[0193] For a logical group G = {I1, I2, ..., I...} containing M images m} Calculate the mean cosine similarity of all image pairs within the group:

[0194] ;

[0195] Where: M×(M-1) / 2 is the number of all possible image pairs in the group; coefficient 2 / (M×(M-1)) is used for normalization;

[0196] Implementation example:

[0197] Suppose the logical group contains 3 images, and the cosine similarity between each pair of images is as follows:

[0198] CosineSim(I1, I2) = 0.95;

[0199] CosineSim(I1, I3) = 0.88;

[0200] CosineSim(I2, I3) = 0.92;

[0201] The team maintains a consistent style as follows:

[0202] StyleUnity(G) = (2 / (3×2)) × (0.95 + 0.88 + 0.92);

[0203] = (1 / 3) × 2.75;

[0204] ≈ 0.917.

[0205] Step 3.3.4: Image filtering and replacement;

[0206] The system sets the style unity threshold T_unity (such as 0.85). During the grouping process:

[0207] If StyleUnity(G) ≥ T_unity, accept this combination;

[0208] If StyleUnity(G) < T_unity, identify the image with the lowest similarity to other images in the group and select an image with a closer style from the candidate pool for replacement;

[0209] The replacement strategy adopts a step-by-step optimization method:

[0210] Calculate the average similarity of each image I i to other images in the group;

[0211] Find the image I with the lowest average similarity min ;

[0212] Select the image I with the highest average similarity to the remaining images in the group from the candidate pool replace ;

[0213] Replace I replace with I min ;

[0214] Recalculate StyleUnity(G) and repeat the above steps until the threshold requirement is met.

[0215] User-defined logic: The user can specify the sorting rule, for example, "Prioritize displaying images with the emotion of 'joy' and sort them in descending order according to the 'vintage' degree of the style".

[0216] Output: One or more sequences of sorted or grouped images.

[0217] In practice, the narrative logic conflicts with the spatial layout. Images that are required to appear sequentially logically may appear in remote fragmented positions in space, which is likely to cause reading fragmentation. This problem is solved by adding narrative adjacency constraints. The following term is introduced in the layout optimization function, that is, to constrain images that are logically adjacent to be as close as possible in the spatial layout:

[0218] .

[0219] Among them, Pairs represents the set of all logically adjacent image pairs ([[]] i, j ) determined in step 3; Pos (i) and Pos(j) respectively represent image i and image jThe center coordinates of the fragment regions assigned in the final composite image layout; ||·||² represents the Euclidean distance. By minimizing L adj This can encourage images with narrative continuity to move closer to each other in space.

[0220] Step 4: Fragmentation and reassembly based on semantic matching;

[0221] The sorted image sequence is applied to a preset fragmentation template to generate the final composite image.

[0222] Fragmented Template: A digital canvas that defines the final image layout, divided into multiple independent fragmented regions. Each fragmented region not only has attributes of position, shape, and size, but more importantly, it carries one or more "semantic anchors." For example, one fragmented region in the template might be anchored as {dimension: 'time', value: 'past'}, while another fragmented region might be anchored as {dimension: 'emotion', value: 'joy'}.

[0223] Semantic matching and filling: Traverse each fragment region in the template. For a given fragment region, the system searches for the image that best matches its semantics in the image sequence generated in step 3, based on its "semantic anchor point." Once found, the content of that image (or its most relevant part) is cropped and filled into the fragment region.

[0224] Generation: After all the fragmented areas are filled, the system synthesizes them into a composite image that is visually broken, but all the fragments are connected by an internal semantic logic.

[0225] In practical implementation, this invention found that when image fragments need to be filled into non-rectangular areas (such as triangular or polygonal areas), common bilinear interpolation will cause image distortion, especially in face and text areas.

[0226] This invention proposes an affine transformation algorithm based on local perception constraints. During the cropping process, deformation penalty weights are added to key points (such as facial feature localization points) to prioritize maintaining the proportion of key regions while allowing slight compression of secondary regions. Optionally, the deformation penalty weight λ_deform ranges from [0.1, 1.0], preferably [0.2, 0.6], and most preferably 0.2 as the default value.

[0227] In a more preferred embodiment, λ_deform can be based on an image quality score. Q base (I) Dynamic adjustments are made, and the specific calculation formula is as follows:

[0228] λ_deform(I) = 0.2 + 0.6 · (1 - Q base (I) );

[0229] This dynamic strategy enables the system to automatically adapt to input images of varying quality, minimizing the negative impact on recall while maintaining high accuracy. Experiments show that, compared to a fixed-weight scheme, this dynamic strategy can improve the overall performance metric (F1-Score) by approximately 7-10 percentage points.

[0230] The slightly compressed image meets one or a combination of the following conditions:

[0231] 1. The compression ratio is between 2 and 10, preferably between 3 and 6, and more preferably about 4;

[0232] 2. The peak signal-to-noise ratio (PSNR) is not less than 35 dB, preferably not less than 38 dB, and more preferably not less than 40 dB;

[0233] 3. The structural similarity index (SSIM) is not less than 0.95, preferably not less than 0.97, and more preferably not less than 0.98;

[0234] 4. When using JPEG encoding, the quality factor (QF) is between 75 and 95, preferably between 85 and 90;

[0235] 5. When using H.264 or H.265 encoding, the quantization parameter (QP) is between 18 and 28, preferably between 22 and 26.

[0236] By constraining the above quantitative indicators, we can ensure that the compressed image achieves effective data reduction while maintaining high visual quality.

[0237] Furthermore, for the same "semantic anchor point," multiple candidate images may appear, making it difficult to select the optimal one. This invention proposes a multi-objective scoring function:

[0238] ;

[0239] in For candidate image features, w is the ideal vector required for the anchor point. style For style dimension, w emotion This is the emotional dimension; parameters α, β, and γ represent the weights adjusted by the user. The system automatically selects the image with the highest score.

[0240] To implement the above method, the present invention also provides a system, the architecture of which is as follows: Figure 2As shown. This system can be used as a standalone software program or integrated into an existing image processing platform, and includes the following functional modules:

[0241] Input module: Responsible for receiving image files uploaded by users and text topics entered by users.

[0242] Semantic analysis engine: The core processing unit of the system, which includes:

[0243] The Natural Language Processing submodule is used to parse core topics. It can employ Transformer-based models (such as BERT) for deep semantic understanding.

[0244] The image recognition submodule integrates multiple deep learning models for content, style, sentiment, and time analysis. These models can be built using open-source frameworks (such as TensorFlow and PyTorch) and pre-trained weights (such as weights trained on ImageNet and COCO datasets).

[0245] The Logic and Sorting module stores and executes logical relationship rules. It receives labeled image data from the semantic analysis engine and logical rules from the user or system, and outputs a sorted sequence of images. Logical relationships are not limited to preset rules but can be automatically learned and generated from a large number of artworks or narrative texts through machine learning models.

[0246] Template Library and Matching Module: This module contains one or more selectable fragmented templates. Its core function is to perform a semantic matching algorithm, precisely mapping the sorted image to each fragment region of the template. The fragmented templates are not static but can be dynamically generated based on the number of input images and their semantic relationships, enabling a more adaptive layout.

[0247] Image generation and output module: responsible for cropping, transforming and finally compositing images based on the matching results, generating high-resolution composite images, and presenting them to the user or saving them as files.

[0248] User interaction interface: (optional, but essential for the application) Provides a graphical interface that allows users to upload data, select or customize logical relationships and fragmented templates, and preview and adjust the final results.

[0249] During testing, it was found that the model inference time was between 10 and 15 seconds, which was a significant performance bottleneck for high-resolution images. In system integration, ONNX (Open Neural Network Exchange) format quantization inference and batch cropping preprocessing were adopted, improving performance by approximately 35%.

[0250] In practical applications, there may be situations where no image in the resource library perfectly matches a certain semantic anchor point. To address this issue, this invention introduces a multi-level fallback strategy to ensure the robustness of the generation process. Through this strategy, the system can still produce a relatively complete work both logically and visually, even when faced with imperfect input. This strategy is divided into five levels (levels 0-4) from high to low priority, as follows:

[0251] 1. Optimal path (level 0): When the face detection confidence is greater than 0.95 and the similarity of the retrieved matching face is greater than 0.85, execute the complete style sentiment transfer process;

[0252] 2. Downgraded Matching (Level 1): When the similarity of matched faces is between 0.6 and 0.85, a Top-K weighted fusion strategy is adopted, and the generation intensity parameters are reduced. The generation intensity parameters include: style transfer intensity coefficient, emotion rendering weight, and feature fusion ratio. The style transfer intensity coefficient ranges from 0.3 to 0.7, and is reduced to 0.5 by default; the emotion rendering weight ranges from 0.2 to 0.6, and is reduced to 0.4 by default; the feature fusion ratio ranges from 0.4 to 0.8, and is reduced to 0.6 by default.

[0253] 3. Generalized Templates (Level 2): ​​When retrieval fails but face detection succeeds, a predefined style / emotion template library is used;

[0254] 4. Minimal Intervention (Level 3): When the generation module fails, it reverts to traditional image processing methods; the traditional image processing methods include one or more of the following: histogram equalization, color space conversion and correction, bilateral filtering denoising, Laplacian sharpening, affine transformation and perspective correction, and image stitching based on Poisson fusion;

[0255] 5. Output as is (Level 4): When the input image cannot be processed, return the original image with additional diagnostic information.

[0256] Whether to trigger a rollback is determined by monitoring thresholds for key quality indicators (such as confidence level, similarity, NIQE score, etc.) at each level. Preferably, the system records the triggering reason, execution level, and result quality each time a rollback occurs, for subsequent analysis and optimization.

[0257] Step S5: Adaptive cropping, transformation, and generative synthesis;

[0258] This step will match the image obtained in step S4. It is precisely filled into the corresponding fragment region F_j.

[0259] 5.1 Optimal Viewpoint Cropping: First, the image is located using a saliency detection model. The most information-rich area is selected, and then, based on the aspect ratio of the fragment F_j, a viewport that maximizes the inclusion of salient content is chosen for cropping.

[0260] 5.2 Locally Perceptive Geometric Transformation: When mapping the cropped rectangular viewport to a fragmented region F_j that may be non-rectangular, a locally perceptive constraint transformation algorithm is employed. This algorithm, when performing affine or perspective transformations, applies a deformation penalty weight to key feature points within the image (such as facial features), prioritizing the preservation of the proportions and shapes of these key regions while allowing secondary background regions to undergo more deformation.

[0261] 5.3 Boundary Processing and Generative Fill: After all fragments are filled, the seams between the fragments are processed. Options include preserving sharp boundaries to enhance the style, or performing edge blending for a smooth transition. Specifically, when the cropped image cannot completely cover the fragmented area, the generative fill module is enabled. This module utilizes techniques such as diffusion models or GANs to intelligently generate natural and harmonious new pixels based on the content of the surrounding filled pixels, seamlessly filling any blank areas.

[0262] 5.4 Global Style Coordination: Finally, a global color and lighting analysis is performed on all filled fragments. Fragments whose hue and contrast deviate too much from the overall atmosphere are automatically corrected to ensure the visual unity of the final composite image.

[0263] Specifically, when the selected image content, even after cropping, cannot completely or naturally fill the target fragment region (especially irregularly shaped regions), or to create a smoother transition between images, this invention can further employ generative filling techniques. This technique utilizes the image inpainting or outpainting capabilities of Generative Adversarial Networks (GANs) or Diffusion Models. Specifically:

[0264] Use the already filled image content as context input.

[0265] Define the blank area to be filled as a mask.

[0266] The generative model is guided to generate natural and harmonious new pixel content within the masked area based on the semantics of the context (such as sky and grass) and style (such as hue and brushstroke).

[0267] This approach not only solved the thorny problem of mismatch between materials and templates, but also greatly enhanced the artistic integrity and visual effect of the final composite image.

[0268] The image fragmentation reconstruction method, or system, with multidimensional semantic association provided by this invention has the following main innovations:

[0269] Multidimensional semantic-driven image association: The core innovation of this invention lies in elevating image association from discrete labels to a continuous high-dimensional semantic space, realizing the quantification, computability, and comparability of the inherent meaning of images. This is the mathematical foundation for achieving truly intelligent reorganization. Image association has been elevated from the traditional visual level (color, shape) to a multidimensional semantic level.

[0270] Deep integration of narrative logic and visual expression: By formalizing the abstract narrative structure into a weighted target semantic vector and a global allocation optimization problem, this invention creatively transforms the process of "storytelling" into a solvable mathematical model, ensuring the logical consistency and narrative expressiveness of the final visual presentation.

[0271] Automated Generation of Visual Works with Complex Meanings: This invention transforms the complex image storytelling process, which requires high levels of creativity and expertise, into a standardized, automated workflow. It not only automates the process but also solves the thorny problem of mismatched materials and templates in real-world applications through multi-level fallback strategies and generative infill technology, enabling the system to produce high-quality works even with imperfect input.

[0272] The image fragmentation reconstruction method, or system, with multidimensional semantic association provided by this invention has the following main advantages:

[0273] Significantly enhances the narrative power of images: The composite images generated by this invention can simultaneously present the passage of time, multiple perspectives, emotional evolution, or different aspects of things within a single frame. Their information density and narrative depth far exceed any single image or traditional jigsaw puzzle. It achieves a leap from "recording a moment" to "telling a story."

[0274] Significantly improves creative efficiency and automation: The complex creative process that originally required hours or even days of manual operation is shortened to a few minutes of automated processing. Users only need to provide materials and high-level creative intentions (theme and logic), and the system can complete all the technical work.

[0275] Pioneering a new paradigm of visual art: "Visual fragmentation under semantic association" is itself a novel art style. This visual effect of "dispersed form but unified spirit" not only has a modern aesthetic but also stimulates viewers' desire for thought and interpretation, providing a brand-new expressive tool for fields such as digital art, advertising design, news reporting, and educational displays.

[0276] High flexibility and scalability: Because both the logical relationships and templates are customizable, the method of this invention possesses extremely high flexibility, adapting to various creative needs. Whether it's personalized artistic creation or commercial content production, suitable application scenarios can be found. Compared to the fixed functions of traditional tools, this invention offers significantly greater scalability.

[0277] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications and improvements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be defined by the claims.

Claims

1. A method for image fragmentation and reconstruction with multidimensional semantic association, characterized in that, Includes the following steps: S1. Receive the image file uploaded by the user and the text topic entered, and parse the core topic; S2. For each input image, use a deep learning model to perform content, style, sentiment, and time analysis, and label it with a set of multi-dimensional semantic tags. S3. Based on the core themes identified, group and sort the labeled image set according to logical relationships; S4. Provide fragmented templates, perform fragmented recombination based on semantic matching, and map the sorted image to each fragment region of the template; S5. Based on the matching results, crop, transform, and synthesize the images to generate the final composite image; S1 performs natural language processing on the core topic text, extracting keywords, topic concepts, and potential narrative structures, and introduces the following context-aware semantic weight matrix: ； t represents the target word, c represents the cultural / regional context, s represents the complete sentence, α, β, and γ are weight parameters learned from training data, BERT(t) represents the word vectors of the pre-trained model, LocalDict(t,c) is a localized dictionary used to map the target word t to a more accurate base word based on the context c, and ContextEmb(t,s) is the context embedding vector generated by analyzing the role of word t in the whole sentence s. S4 includes: Each fragmented template is divided into multiple independent fragment regions; Each fragment region has location, shape, and size attributes, and has at least one or more semantic anchors; For a certain fragment region, the image that best matches its semantics is found in the image set generated in step S3 based on its semantic anchor point; after finding it, the content of the image is cropped and filled into the fragment region. In step S4, when multiple candidate images appear for the same semantic anchor point, the following multi-objective scoring function is designed, and the image with the highest score is automatically selected: ； in For candidate image features, w is the ideal vector required for the anchor point. style For style dimension, w emotion It is the emotional dimension, and the parameters α, β, and γ represent the weights that the user adjusts.

2. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 1, characterized in that, S2 represents each image as a semantic vector containing multiple key-value pairs, including: Content element annotation: Identify key objects, scenes, and activities in images using object detection and image classification models; Style feature annotation: using convolutional neural networks to classify the artistic style of images; Sentiment annotation: Using a model trained on a sentiment image dataset, analyze the overall emotional atmosphere conveyed by the image; Time attribute annotation: Extract the shooting time from the image's metadata, or infer its relative time attribute through image content analysis.

3. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 1, characterized in that, The S2 quickly assesses the basic image quality score from the dimensions of sharpness, noise, and contrast. Q base (I) : Q base (I) = w_s × S(I) + w_n × N(I) + w_c × C(I); Where w_s, w_n, w_c are weighting coefficients, and their sum is 1; S(I) is the normalized sharpness score; N(I) is the normalized noise score; and C(I) is the normalized contrast score. when Q base (I) Images below a preset threshold are identified as low-quality or blurry, and further fine-tuned using a deep learning quality evaluation function Q(I); S2 employs the following adaptive confidence calibration algorithm when processing low-quality or blurry images: ； Where: Q(I) is the image quality evaluation function: BRISQUE is a blind / no-reference image spatial quality evaluator. Here, T(age) is the natural image quality evaluator, and T(age) is the time decay correction factor. , where age represents the year of the image.

4. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 1, characterized in that, S3 uses the following logical relationship for grouping and sorting: Multidimensional deconstruction of the same thing: Group images with the same core content element labels together, and then sort them according to style, time or sentiment dimensions; maximize the cosine similarity between the style feature vectors of all images in the same group to improve style uniformity; Continuous narrative: Linearly sort images according to temporal attributes or causal logic of events to construct a story flow; minimize the total variance of the temporal attributes of all images in the story flow so that they are smoothly arranged in the time dimension; Contradictions and Contrasts: Euclidean distance is used to calculate the emotional contrast between two images, and images with opposite or conflicting emotional / content labels are paired or grouped.

5. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 4, characterized in that, S3 further includes introducing the following term into the layout optimization function by adding narrative adjacency constraints to spatial layout conflicts, such that logically adjacent images are spatially adjacent: ； Where, Pos(i) and Pos(j) Representing images respectively i and images j The center coordinates of the fragment region assigned in the final composite image layout.

6. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 1, characterized in that, S4 also includes the following multi-level fallback strategy: Level 0, Optimal Path: When the confidence score of the key content element detection is greater than 0.95 and the semantic similarity of the retrieved matching content is greater than 0.85, the complete style sentiment transfer process is executed; Level 1, Degraded Matching: When the semantic similarity of the matched content is between 0.6 and 0.85, a Top-K weighted fusion strategy is adopted, and the generation intensity parameter is reduced; the generation intensity parameter includes the style transfer intensity coefficient, sentiment rendering weight, and feature fusion ratio; The value range of the style transfer intensity coefficient is 0.3 to 0.7, the value range of the emotion rendering weight is 0.2 to 0.6, and the value range of the feature fusion ratio is 0.4 to 0.

8. Level 2, Generalized Templates: When retrieval fails but key content elements are detected successfully, a predefined style / sentiment template library is used; Level 3, Minimal Intervention: When the generation module fails, it reverts to traditional image processing methods; Level 4, Original Output: When the input image cannot be processed, return the original image with additional diagnostic information.

7. The image fragmentation and reconstruction method with multidimensional semantic association as described in claim 1, characterized in that, S5 includes: S51. Optimal window cropping: First, locate the most informative region in the image using a saliency detection model. Then, combine the aspect ratio of the fragments to select a window that maximizes the inclusion of salient content for cropping. S52. Locally Perceptive Geometric Transformation: When mapping the cropped rectangular viewport to a non-rectangular fragmented region, a locally perceptive constraint transformation algorithm applies a deformation penalty weight to key feature points within the image, prioritizing the preservation of the proportions and shapes of these key regions while allowing secondary background regions to generate more deformation; the deformation penalty weight λ_deform ranges from [0.1, 1.0], or is determined based on the image quality baseline score. Q base (I) Dynamic adjustment is performed, and the calculation formula is: λ_deform(I) = 0.2 + 0.6 · (1 - Q base (I) ); S53, Boundary Processing and Generative Fill: After all fragments are filled, the seams between fragments are processed; when the cropped image cannot completely cover the fragment area, the generative fill module is enabled, using a diffusion model or generative adversarial network to generate new pixels based on the content of the surrounding filled pixels to seamlessly fill the blank area. S54. Global Style Coordination: Finally, a global color and lighting analysis is performed on all filled fragments, and fragments whose hue and contrast deviate from the overall atmosphere by more than a preset threshold are automatically corrected.

Images