Cross-scene special effect style adaptation method based on meta learning

Abstract

The present application relates to the technical field of image processing and style transfer, and specifically discloses a cross-scene special effect style self-adaptive method based on meta-learning, which constructs an implicit high-dimensional meta style prototype vector library offline, dynamically selects and fuses the prototype vector according to the input scene semantic description in real time, and enhances the universality and flexibility of style transfer; scene analysis and saliency detection are used to differentially process different regions with different style transformation tolerance, thereby reducing the risk of distortion of key content; a human-computer interaction interface is introduced to balance automation and subjective setting by combining user subjective style preference with scene objective analysis results, thereby improving the controllability of the results; finally, multi-scale features are extracted through a multi-layer convolutional network, linear affine adjustment is performed in combination with a differential style field, and a target picture that maintains content recognition and is rich in artistic style is rendered. The present application improves the adaptability and effect quality of style transfer.

Description

Technical Field

[0001] This invention relates to the fields of image processing and style transfer technology, and more specifically, to a cross-scene special effects style adaptation method based on meta-learning. Background Technology

[0002] In current image style transfer applications, applying a specific artistic style uniformly to visual scenes with different content is a common requirement. However, since different visual scenes often contain significantly different semantic content, layout structures, and visual focal points, using fixed style parameters or a single style model may lead to adaptation issues. For example, for portrait scenes and natural landscape scenes, the same set of style parameters may result in distortion of facial details in the portrait or excessive distortion of landscape colors. This constitutes a technical challenge: how to enable style transfer methods to adaptively adjust the intensity, method, and emphasis of style application according to the specific content of different scenes, so as to achieve a harmonious integration of style effects and scene content, while maintaining good content recognizability.

[0003] Patent CN119599860A discloses a method and system for style transfer of low-quality aircraft infrared images across illumination domains. The method includes steps such as image preprocessing, building an image style transfer model, training the image style transfer model, and image style transfer. Based on a preprocessing strategy and by building a generative network based on multi-scale feature fusion and attention mechanisms in the image style transfer model, it effectively achieves style transfer for low-quality aircraft infrared images with poor resolution, low contrast, and low signal-to-noise ratio. This solves the problems of target variation, background clutter, and loss of detail that lead to unusable images in cross-illumination domain style transfer of low-quality aircraft infrared images.

[0004] Patent CN120953492A discloses a method, computer device, readable storage medium, and program product for cross-domain transfer of 3D scenes based on neural radiation field representation. The method includes: obtaining a real image for 3D scene reconstruction; inputting 3D scene information into a geometry module to extract content features; inputting 3D scene information into a source domain shading module to extract source domain style features; volume rendering to obtain a first rendered image; training and optimizing the geometry module and the source domain shading module until expectations are met, then freezing the geometry module; inputting 3D scene information into a target domain shading module to extract target domain style features; volume rendering to obtain a second rendered image; performing style transfer on the real image to obtain a stylized image; training and optimizing the target domain shading module using contrastive loss until expectations are met, then freezing the target domain shading module; receiving a specified viewpoint direction; using the geometry module and the target domain shading module to output features respectively; and obtaining the target domain image through volume rendering.

[0005] Existing technologies have shortcomings in addressing the complexity and diversity of cross-scene style transfer: First, these methods typically require independent, time-consuming end-to-end model training for each specific "source-target" image set (e.g., day and night, specific artistic styles). The learned style representation is rigidly fixed in a single network weight, resulting in poor generalization ability. Furthermore, when faced with new scenes or styles, data must be collected again and the training process restarted, making it difficult to achieve zero-shot or few-shot generalization capabilities for "one-time training, multi-scene adaptation." Second, in terms of the granularity of style application, existing technologies often employ globally consistent style mapping logic, lacking differentiated perception of the visual importance and style tolerance of different semantic entities within the input image (e.g., foreground figures and background buildings). This easily leads to sacrificing the detail fidelity of key content (e.g., facial features) when pursuing overall style transfer, failing to achieve "content-sensitive" local adaptive style modulation. Finally, these rigid model architectures limit the real-time intervention of human subjective aesthetic intentions during the reasoning stage. The system cannot dynamically adjust the output effect according to the user's immediate preferences (such as fine-tuning of style intensity), resulting in a single stylized result and a lack of controllability. Summary of the Invention

[0006] In view of this, in order to solve the problems mentioned in the background technology, a cross-scene special effects style adaptation method based on meta-learning is proposed.

[0007] The objective of this invention can be achieved through the following technical solution: This invention provides a cross-scene special effects style adaptation method based on meta-learning, including the following steps: S1, acquire style feature asset data, construct a prototype library containing implicit high-dimensional meta-style prototype vectors in memory, load a pre-trained style field generation network, establish a communication link between the prototype library and the style field generation network, and generate a style adaptation system.

[0008] S2. Receive the current image frame to be processed, parse the pixel structure and dependency relationship of the current image frame to be processed, extract multi-dimensional scene perception data, synthesize and output a customized structured semantic descriptor.

[0009] S3. Import the customized structured semantic descriptor into the style adaptation system, perform forward mapping and similarity calculation through the style field generation network, and generate addressing weights pointing to the implicit high-dimensional style prototype vector.

[0010] S4. Obtain the human subjective style intensity configuration parameters, and superimpose the human subjective style intensity configuration parameters with the addressing weights to generate the modulated addressing weights.

[0011] S5. Based on the modulated addressing weights, extract the corresponding implicit high-dimensional style prototype vectors from the prototype library, and synthesize a differential style field for all pixels of the current image frame through spatial weighted interpolation.

[0012] S6. Read the differential style field and the current image frame to be processed, extract the multi-scale feature representation base map through a multi-layer convolutional network, use the differential style field to perform linear affine adjustment on the multi-scale feature representation base map, and inversely render the target stylized image.

[0013] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) The present invention constructs a set of implicit high-dimensional meta-style prototype vectors offline using a meta-learning algorithm to form a prototype library. These prototype vectors represent transferable latent style preferences independent of specific image content. During real-time style transfer, the system can dynamically select and fuse multiple prototype vectors from the prototype library according to the semantic description of the input scene, rather than relying on a single fixed style parameter. This mechanism enables the basic parameter library of style transformation to have cross-scene universality and combinatorial flexibility, thereby enhancing the adaptability of the style transfer method to different visual scene content and avoiding the need to retrain a complete model for each new scene.

[0014] (2) This invention analyzes the content focus of the input image through a scene analysis step using a saliency detection algorithm, segments and defines regions with different tolerances to style transformation, and generates a content sensitivity mask. In the subsequent style application process, this mask guides the system to allocate lower style transformation intensities to highly sensitive regions and higher style transformation intensities to less sensitive regions. This mechanism of differentiated processing based on content importance reduces the probability of excessive distortion or information loss in key content areas of the image during the stylization process, improving the content fidelity and subjective visual experience of the final result while ensuring the artistic stylization effect.

[0015] (3) This invention establishes a weighted addressing mechanism based on scene structured semantic description. The system first parses the semantic information of the input image and generates a customized structured semantic descriptor. Then, it uses a style field generation network to calculate a set of addressing weights pointing to various prototypes in the prototype library based on the descriptor. These addressing weights are spatially non-uniformly distributed, meaning that pixels at different locations in the image may correspond to different prototype combination weights. This mechanism enables the system to dynamically generate a non-uniformly distributed differential style field based on the semantic features of local regions of the image. This achieves differentiated and refined control of style application in the image space, allowing the style effect to better fit the local scene content.

[0016] (4) This invention introduces a human-computer interaction interface to receive global style intensity configuration parameters provided by the user. The system combines this subjective parameter with a region scaling compensation factor automatically calculated based on scene analysis to modulate the original addressing weights. This modulation process can integrate the user's subjective style preferences with the system's analysis results of the objective structure of the scene. This mechanism ensures that the final output stylization effect is no longer a simple automated result or a simple subjective setting result, but achieves a balance between the two, realizing adaptive adjustment between user input and actual scene perception needs, and enhancing the controllability and adaptability of the results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method steps of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides a cross-scene special effects style adaptation method based on meta-learning, including: S1, acquiring style feature asset data, constructing a prototype library containing implicit high-dimensional meta-style prototype vectors in memory, loading a pre-trained style field generation network, establishing a communication link between the prototype library and the style field generation network, and generating a style adaptation system.

[0021] In a specific embodiment of the present invention, a communication link is established between the prototype library and the style field generation network to generate a style adaptive system, including: constructing a prototype library containing multiple sets of implicit high-dimensional style prototype vectors in a solid-state medium storage space, wherein the implicit high-dimensional style prototype vectors represent potential universal style transformation preference parameters independent of a single basic art filter framework.

[0022] The style field generation network, which has been pre-trained offline, is loaded into the central processing unit's running memory. The style field generation network serves as a data processing engine that establishes a feature addressing and solving mechanism for the corresponding prototype library resources.

[0023] A data reading and instruction exchange communication link is established between the prototype library and the style field generation network on the underlying hardware motherboard bus to generate a style adaptive system in a memory-resident, ready-to-go state.

[0024] Specifically, this step begins with the CPU of the execution terminal accessing the local solid-state storage space according to a preset startup instruction, retrieving and locating a style feature asset data file generated and packaged by the meta-learning algorithm during the offline training phase. The processor reads the style feature asset data file into the system's runtime memory in binary stream form via the file system interface. The style feature asset data file contains two core data parts. The first part consists of multiple sets of implicit high-dimensional meta-style prototype vectors. The system allocates a dedicated high-speed cache area in memory, deserializes these vector data, loads them into this area, and constructs a hash mapping table with integer indices as primary keys and vector data as values, forming a prototype library for fast random access in subsequent steps. The second part of the data consists of the pre-trained style field generation network structure definition and weight parameters. The processor calls the application programming interface of a deep learning inference framework, such as TensorFlow Lite or ONNX Runtime, to instantiate the network computation graph in runtime memory according to the network structure definition, and fills the weight parameters into the corresponding network layer nodes, completing the loading of the style field generation network. Finally, the operating system's memory manager allocates virtual memory addresses for the loaded prototype library and style field generation network. The execution engine of the deep learning inference framework then establishes direct access pointers to the prototype library's memory addresses, thus logically forming a high-speed data reading and instruction exchange communication link between the prototype library and the style field generation network based on a memory bus. After this process is completed, the entire system, including the data and model, resides in the runtime memory as a unified functional entity and enters a suspended standby state, forming a memory-resident, standby, ready style adaptive system, awaiting the triggering of subsequent processing tasks.

[0025] The implicit high-dimensional meta-style prototype vector is a mathematical construct that does not directly correspond to any specific visual effect, but rather represents a basic style transformation trend or preference in a high-dimensional latent space. Specifically, it is a vector composed of N-dimensional floating-point numbers. This vector is obtained through an offline meta-learning training process, aiming to capture commonalities and transferable style features across various art styles. In this embodiment, the value of N is set to a range of 256 to 512, representing an experimental result that balances the richness of style representation with computational resource consumption. The prototype library is a structured dataset; in this embodiment, it contains 64 independent implicit high-dimensional meta-style prototype vectors. This number is set based on a balance between achieving diverse style synthesis and memory usage on typical consumer devices. The style field generation network is a lightweight feedforward neural network, such as employing deep separable convolutions or the MobileNet architecture. Its design goal is to receive structured information about the scene and quickly compute the weights used to guide the combination of implicit high-dimensional meta-style prototype vectors. A style-adaptive system refers to a software system state in which the data entities of the prototype library and the computational entities of the style field generation network are uniformly managed and scheduled in the runtime memory.

[0026] For example, suppose the style feature asset data is stored as a binary file named `style_assets.bin` on the solid-state drive of the execution terminal. Upon system startup, the CPU first reads this file. The first 8KB of data is a prototype library containing 64 prototype vectors, each consisting of 256 32-bit floating-point numbers. The processor reads this 8KB of data all at once into memory at address 0x1000A000 and creates an index array where the i-th element points to the starting address of the i-th prototype vector. Next, the processor reads the remaining 5.4MB of data from the 8KB offset in the `style_assets.bin` file; this data is a style field generation network model in ONNX format. The processor calls the ONNXRuntime API functions to load and parse this 5.4MB of model data, constructing the corresponding computation graph and parameter nodes in memory. After this operation, the prototype library data is located at address 0x1000A000, and an instance of the style field generation network is created and ready. The network instance maintains a data pointer to address 0x1000A000, which is used to directly extract the implicit high-dimensional style prototype vector from memory during subsequent inference computations. At this point, the style adaptation system is deployed, occupying approximately 5.41MB of runtime memory, and enters a waiting state, ready to receive the next input image frame for processing.

[0027] S2. Receive the current image frame to be processed, parse the pixel structure and dependency relationship of the current image frame to be processed, extract multi-dimensional scene perception data, synthesize and output a customized structured semantic descriptor.

[0028] In a specific embodiment of the present invention, the pixel structure and dependencies of the current image frame to be processed are analyzed, multi-dimensional scene perception data is extracted, and a customized structured semantic descriptor is synthesized and output, including: activating the edge detection convolution operator to detect the pixel transition smoothing level within the current image frame to be processed, and generating a semantic gradient map in the spatial dimension.

[0029] The saliency detection algorithm is run to evaluate the content focus of the current image frame to be processed, segment and define the preserved region and the transformed region, and output the content sensitivity mask.

[0030] In a specific embodiment of the present invention, the output content sensitivity mask includes: controlling the central processing unit to run a saliency detection algorithm to calculate the content focus score for each pixel position of the current image frame to be processed.

[0031] The content focus score is compared pixel by pixel with the preset safety tolerance score.

[0032] Pixels with content focus scores higher than the safety tolerance score are assigned to the reserved area, while pixels with content focus scores lower than the safety tolerance score are assigned to the transform area, thereby generating a binarized content sensitivity mask.

[0033] Measure the numerical distances of semantic objects in the two-dimensional coordinate plane and match the spatial binding tightness to construct a region association map.

[0034] In a specific embodiment of the present invention, measuring the arrangement distance of semantic objects in a two-dimensional coordinate plane and matching the spatial binding tightness to construct a region association map includes: measuring the arrangement distance of semantic objects with different structural attributes in a two-dimensional coordinate plane using the Euclidean distance calculation equation.

[0035] Retrieve a pre-defined table of co-occurrence probabilities for each category and match the spatial density of each region.

[0036] The numerical values ​​of the arrangement distance are fused with the spatial density and encoded to construct a regional association map.

[0037] Based on the feature depth dimension, the semantic gradient map, content sensitivity mask and region association map are connected to generate a customized structured semantic descriptor.

[0038] Specifically, this step begins when the style adaptation system receives an external interrupt signal indicating that the video acquisition hardware is ready to provide new image data. The system first activates the camera sensor via the graphics driver interface, instructing it to capture a continuous real-time augmented reality video stream at a preset frame rate, such as 30 frames per second. For each frame in the video stream, the system uses a vertical sync signal as the time reference for inter-frame synchronization, capturing the currently displayed image on the screen at the moment the signal is triggered and treating it as an independent current image frame to be processed. This frame data, such as a 1920x1080 pixel 32-bit RGBA format bitmap, is immediately copied from main memory to the graphics processor's dedicated video memory frame buffer to reduce memory access latency for subsequent calculations. Next, the graphics processor initiates a parallel computing task, loading a pre-compiled edge detection convolution operator. This operator performs a two-dimensional convolution operation on the luminance channel of the current image frame to detect the rate of change in pixel intensity at the boundaries of different target semantic object regions, i.e., the pixel transition smoothing level. Following this, a computational core based on a second-order reciprocal algorithm, such as the Laplacian operator, is used to process the convolution results, extracting and quantifying these changes to generate a single-channel floating-point matrix of the same size as the original image—a semantic gradient map in spatial dimensions. Simultaneously, another computational thread is scheduled to the central processing unit or a dedicated AI coprocessor to activate a lightweight saliency detection algorithm that performs a full-image analysis of the current image frame. This algorithm calculates a content focus score for each pixel location in the image. The system compares these scores pixel-by-pixel with a preset safety tolerance score, assigning pixels with scores higher than the tolerance score to a first reserved region and those lower to a second transformed region, thereby generating a binary content sensitivity mask. Furthermore, the system runs a lightweight semantic segmentation network to identify the main object categories and their pixel regions in the image. For each identified semantic object region, the system calculates the two-dimensional coordinates of its geometric center point and uses the Euclidean distance equation to calculate the numerical distance between the center points of different object regions. Simultaneously, the system accesses a pre-defined class co-occurrence probability table, which stores the frequency statistics of different object categories appearing simultaneously in the training dataset. The system queries this table to match the spatial binding density between regions in the current frame and constructs a region association map by combining distance and density information. Finally, a tensor concatenation operator is invoked. This operator merges semantic gradient maps, content sensitivity masks, and region association maps with the same spatial dimension in a stacked manner along the feature depth dimension to form a multi-channel tensor. This tensor is the final output, a customized structured semantic descriptor carrying multi-dimensional scene-aware data parameters.

[0039]

[0040] The above formula is the Euclidean distance calculation equation, used to calculate the numerical value of the arrangement distance. Wherein, Indicates the first The semantic object and the first Euclidean distance between semantic objects. Representing the Two-dimensional pixel coordinates of the geometric center of a semantic object. Representing the The formula represents the two-dimensional pixel coordinates of the geometric center of a semantic object. The unit of measurement is pixel length, which remains consistent throughout the entire computational system and requires no normalization.

[0041] The current image frame to be processed is a digital image with a specific resolution, such as 1920x1080 pixels, using the RGBA8888 color format. The edge detection convolution operator is a 3x3 or 5x5 convolution kernel, such as a Laplacian kernel, where the center element is positive, surrounding elements are negative, and the sum of all elements is zero. The semantic gradient map is a single-channel floating-point tensor, with each element ranging from 0.0 to 1.0, representing the edge strength at the corresponding pixel location. The safety tolerance score is a preset floating-point threshold, ranging from 0.6 to 0.8. In this embodiment, the value is set to 0.7, an empirical value obtained statistically based on saliency analysis of a large-scale image dataset to balance content fidelity and artistic stylization. The content sensitivity mask is a single-channel integer tensor, where an element value of 1 indicates that the corresponding pixel belongs to the first reserved region that should be protected, and a value of 0 indicates that it belongs to the second transformation region where a larger style transformation can be performed. The category co-occurrence probability table is a static data structure stored as a hash table, where the keys are object category pairs and the values ​​are probability values ​​between 0.0 and 1.0. The region association graph is a multi-channel floating-point tensor, the number of channels of which depends on the dimensions of the association information to be encoded, such as distance and category co-occurrence probability. The customized structured semantic descriptor is the final multi-channel tensor generated in this step, used as input to the style field generation network in the next step.

[0042] For example, the style adaptation system receives a 1920x1080 pixel current image frame to be processed, depicting a street scene containing a person, a car, and distant buildings. First, a 3x3 Laplacian operator convolves the image frame on the GPU, generating a 1920x1080x1 semantic gradient map, where pixel values ​​at the outlines of the person and car are close to 1.0, while pixel values ​​at flat building walls and sky areas are close to 0.0. Next, a saliency detection network runs on the CPU, outputting a saliency map where the saliency score for the person's face is 0.9, for the car license plate area is 0.85, and for most building walls is 0.3. The system uses a preset safety tolerance score of 0.7 for comparison, generating a 1920x1080x1 content sensitivity mask, where pixel values ​​for the person's face and license plate areas are 1, and for other areas are 0. Subsequently, semantic segmentation identified the center of the person region R1 as (400, 500) and the center of the car region R2 as (1000, 800). The system then calculated the Euclidean distance between them. for This is approximately 670.8 pixels. The system simultaneously queries a category co-occurrence probability table, obtaining a co-occurrence probability of 0.65 for P(person, car). These values ​​are encoded into a 1920x1080x2 region association map. Finally, a tensor concatenation operator merges the aforementioned 1920x1080x1 semantic gradient map, 1920x1080x1 content sensitivity mask, and 1920x1080x2 region association map along the channel dimension, generating a 1920x1080x4 customized structured semantic descriptor. This descriptor is stored in a specified video memory address, awaiting retrieval by the style field generation network.

[0043] S3. Import the customized structured semantic descriptor into the style adaptation system, perform forward mapping and similarity calculation through the style field generation network, and generate addressing weights pointing to the implicit high-dimensional style prototype vector.

[0044] In a specific embodiment of the present invention, forward mapping and similarity calculation are performed through a style field generation network to generate addressing weights pointing to implicit high-dimensional style prototype vectors. This includes: performing fully connected forward mapping propagation on customized structured semantic descriptors in the pre-perception layer of the style field generation network to extract spatial semantic tensors.

[0045] The set of cosine similarity values ​​between the spatial semantic tensor and the corresponding implicit high-dimensional style prototype vectors is calculated by vector multiplication.

[0046] A nonlinear normalization function is introduced to process the set of cosine similarity values ​​and compress it into the standard probability space, generating addressing weights that point to implicit high-dimensional style prototype vectors.

[0047] Specifically, the execution flow of this step is entirely within the closed loop of the style adaptation system, with the style field generation network undertaking the core computational tasks. First, the system reads the customized structured semantic descriptor generated by S2 from the GPU memory and uses it as the input tensor. Without any preprocessing, it is directly fed into the data input node of the style field generation network. The customized structured semantic descriptor then undergoes a fully connected forward mapping propagation in the pre-perception layer of the style field generation network. This pre-perception layer linearly transforms the multi-channel descriptor vector at each spatial location using a weight matrix and applies a non-linear activation function, generating a new tensor with the same dimension as the implicit high-dimensional style prototype vector. This tensor is the spatial semantic tensor. Subsequently, the system constructs a computational grid, or resolution lattice, based on the original resolution of the current image frame to be processed, using each smallest independent unit pixel as the starting point for traversal computation. For each coordinate point on the grid, the system performs the following operations in parallel: extracting the semantic vector corresponding to that coordinate point from the spatial semantic tensor and reading all implicit high-dimensional style prototype vectors from the prototype library. Then, through vector multiplication—specifically, dot product followed by the product of the L2 norms of each vector—the cosine similarity between the semantic vector and each implicit high-dimensional style prototype vector is calculated, thus generating a set of cosine similarity values ​​for that coordinate point. Finally, the system introduces the Softmax nonlinear normalization function to process the cosine similarity value set calculated independently for each coordinate point. This function amplifies the differences between similarities through exponentialization and normalizes the original similarity value set by summing all exponentialization results, compressing and converting it into a standard probability distribution with a sum of 1. The probability distribution set generated across all spatial coordinate points collectively constitutes a tensor with the same size as the original image but a depth equal to the prototype library size, representing the addressing weights that characterize the proportion of features corresponding to all latent prototype set members for each spatial coordinate point in the current image frame to be processed.

[0048]

[0049] The above formula is the equation for calculating cosine similarity. Wherein, Indicates pixel coordinates At that point, the vector of the spatial semantic tensor With the prototype library An implicit high-dimensional style prototype vector The cosine similarity between them. · Represents the vector dot product operation. It represents the L2 norm of a vector.

[0050]

[0051] The above formula is the Softmax nonlinear normalization function. Wherein, In pixel coordinates At that location, the corresponding addressing weight in the generated addressing weight The weight components of each prototype. It is the cosine similarity value corresponding to that point. It represents the total number of prototype vectors in the prototype library. This is a temperature parameter used to adjust the smoothness of the probability distribution. The calculation results in both formulas are dimensionless scalars, so there is no issue of dimension inconsistency.

[0052] The spatial semantic tensor is an intermediate feature tensor with the same spatial resolution as the current image frame being processed, but with a channel count or depth equal to the dimension of the implicit high-dimensional style prototype vector. The resolution lattice is a logical structure used to guide computational traversal, with a resolution strictly equal to the resolution of the current image frame being processed. The cosine similarity value set is a one-dimensional vector at each pixel with a dimension equal to the size of the prototype library, and its element values ​​range from -1.0 to 1.0. The temperature parameter in the Softmax nonlinear normalization function... This is a preset hyperparameter with a value ranging from 0.1 to 1.0. In this embodiment, it is set to 0.5 based on experimental tuning to achieve a balance between highlighting the dominant style and integrating the influence of multiple styles. The addressing weight is the final output of this step, representing a spatial size relative to the original... Figure 1 The number of channels is a floating-point tensor equal to the number of prototype vectors in the prototype library, and the sum of the channel vector elements at each spatial location is 1.

[0053] For example, the system reads a customized 1920x1080x4 structured semantic descriptor generated by S2. Consider the descriptor vector at pixel (200, 300) as [0.8, 0.0, 500.0, 0.4]. This vector is input to the pre-perceptual layer of the style field generation network, which maps it to a 256-dimensional vector as the value of the spatial semantic tensor at position (200, 300), denoted as... Next, the system retrieves the first prototype vector from the prototype library, which contains 64 prototype vectors. and the second prototype vector System calculation By comparing their cosine similarity, we obtain , Assume that the similarities calculated with the remaining 62 prototype vectors constitute a complete set of cosine similarity values ​​[0.85, 0.30, ...]. Then, using... The Softmax function processes this set. It calculates the first weight component: Calculate the second weight component: Assume that, after calculation, the sum of all 64 exponential terms... If it is 12.0, then , This process is performed in parallel on all 1920x1080 pixels, ultimately generating a 1920x1080x64 addressable weight tensor, where the initial value of the 64-channel vector at position (200, 300) is [0.456, 0.152, ...].

[0054] S4. Obtain the human subjective style intensity configuration parameters, and superimpose the human subjective style intensity configuration parameters with the addressing weights to generate the modulated addressing weights.

[0055] In a specific embodiment of the present invention, obtaining the human subjective style intensity configuration parameter, superimposing the human subjective style intensity configuration parameter with the addressing weight to generate the modulated addressing weight, includes: constructing a low-level operating system interrupt response listener to capture intensity change update events triggered by graphical touch interaction components, and extracting the human subjective style intensity configuration parameter.

[0056] Analyze the pixel content sensitivity of the current image frame to be processed, and calculate the region scaling compensation factor with anti-boundary distortion properties.

[0057] The executor multiplies the subjective style intensity configuration parameter with the regional scaling compensation factor to obtain the modulation scalar.

[0058] The modulation scalar is mathematically superimposed into the addressing weight matrix dimension, and the result is normalized in the feature channel dimension to ensure that the sum of the weights is 1, thus generating the modulated addressing weights.

[0059] Specifically, this step begins with the triggering of a low-level operating system interrupt response listener, which specifically monitors state changes of graphical touch interaction components. When the user interacts with the slider control on the human-computer interface, an intensity change update event is generated. The listener captures this event and extracts a dimensionless floating-point number from the event payload; this value is the global style intensity configuration parameter. Simultaneously, a parallel computation task is initiated, which reads the content sensitivity mask generated by S2. By scanning the mask tensor element-wise, the system counts the total number of pixels with a value of 0, which is the area of ​​the second transformation region. This count is then divided by the total number of pixels in the image to calculate the proportion of the second transformation region in the entire frame. Based on this proportion, the system substitutes a preset nonlinear mapping function to calculate a region scaling compensation factor that smooths style transitions and prevents boundary distortion. Subsequently, a hardware multiplier is invoked to multiply the global style intensity configuration parameter and the region scaling compensation factor; the product is the modulation scalar set for the current frame and the current user. Based on the content sensitivity mask, the system broadcasts the modulation scalar to a modulation matrix that has the same spatial and channel dimensions as the addressing weights, but is padded with zeros at the corresponding positions in the first reserved region. Finally, a vector adder performs element-wise mathematical superposition of the addressing weights and the modulation matrix, and renormalizes the result along the channel dimension at each pixel position to ensure that the sum of the weights is 1, thereby generating the final modulated addressing weights that incorporate the user's subjective preferences.

[0060]

[0061] The above formula is used to calculate the regional scaling compensation factor. .in, The number of pixels in the second transform region is obtained by traversing the content sensitivity mask. It represents the total number of pixels in the current image frame to be processed. It is a preset gain coefficient used to control the strength of the compensation. This is a reference scale threshold. When the scale of the transformed region equals this value, the compensation factor is 1, and no scaling effect is produced. This formula is dimensionless.

[0062] The subjective style intensity configuration parameter is typically a floating-point number ranging from 0.0 to 2.0, where 1.0 represents the default intensity, 0.0 represents no stylization effect, and a value greater than 1.0 represents an enhancement effect. This value is provided in real-time by controls such as sliders or knobs in the graphical user interface. The second transformation region is an image region with high tolerance for style transformation, defined by the saliency detection algorithm in S2. The region scaling compensation factor is set based on the fact that when the transformable region is too small or too large, directly applying the global intensity may lead to abrupt style changes in the transformed region or an imbalance in the overall image style. This factor aims to dynamically adjust the actual effect value of the style intensity according to the relative size of the transformable region. Gain coefficient. The value ranges from 0.2 to 0.8, referencing the proportional threshold. The value ranges from 0.4 to 0.6, and these values ​​are empirical values ​​derived from statistical experiments on subjective visual quality assessment of a large number of images from different scenes. In this embodiment, Set to 0.5. Set to 0.5. The modulation scalar is the final intensity adjustment value that combines global intensity preference with scene structure adaptability. The modulated addressing weight is the final output of this step. Its data structure is exactly the same as the addressing weight output by S3, but its numerical distribution has been adjusted by user input and scene analysis results.

[0063] For example, the system first captures the user adjusting the interface slider to a specific position through the underlying operating system interrupt response listener, and extracts the global style intensity configuration parameter as 1.5. Next, the system analyzes the 1920x1080 pixel content sensitivity mask generated by S2. Statistical analysis reveals that there are 207,360 pixels in the first reserved area with a value of 1, and 1,866,240 pixels in the second transformation area with a value of 0. Based on this, the system calculates the occupancy ratio of the second transformation area to be 0.9. Subsequently, the system uses preset parameters... and Substitute the values ​​into the formula to calculate the regional scaling compensation factor, and obtain... Then, the system multiplies the global style intensity configuration parameter 1.5 with the region scaling compensation factor 1.2 to obtain a modulation scalar of 1.8. For the addressing weight vector [0.456, 0.152, ...] at pixel (200, 300) in S3, since this point is located within the second transform region, the system applies the modulation scalar of 1.8 to this vector, for example, by multiplying it with the maximum value component in the vector to enhance it, resulting in a new non-normalized vector [0.8208, 0.152, ...]. Finally, this new vector is re-normalized to obtain the new value of the modulated addressing weight vector at this point. This modulation process covers all pixels in the second transform region, while the addressing weights in the first reserved region remain unchanged, ultimately generating a complete 1920x1080x64 modulated addressing weight tensor.

[0064] S5. Based on the modulated addressing weights, extract the corresponding implicit high-dimensional style prototype vectors from the prototype library, and synthesize a differential style field for all pixels of the current image frame through spatial weighted interpolation.

[0065] In a specific embodiment of the present invention, a differential style field is synthesized for all pixels of the current image frame to be processed through spatial weighted interpolation operation, including: using the non-zero array index of the modulated addressing weight as the retrieval key, and extracting the corresponding implicit high-dimensional style prototype vector in parallel from the prototype library.

[0066] The matrix eigenvalues ​​of the implicit high-dimensional style prototype vector are extracted and combined with the modulated addressing weights to perform spatial weighted interpolation operations, which are then aggregated to generate style transformation feature vectors for individual isolated pixels.

[0067] The style transformation feature vector clusters generated globally are progressively advanced along the row and column directions of the current image frame to be processed and converged to form a differential style field.

[0068] Specifically, the core of this step lies in using the modulated addressing weights generated by S4 to retrieve and fuse data from the prototype library to construct a fine-grained control field guiding the final image rendering. The system first traverses each spatial coordinate point of the modulated addressing weight tensor in parallel. For each coordinate point, the system checks its corresponding 64-dimensional weight vector. Using the non-zero array index of this weight vector as the retrieval key for the parallel query, the system initiates a data read request to the prototype library. As an efficient key-value store, the prototype library, upon receiving these keys, extracts the actual implicit high-dimensional style prototype vector entities corresponding to the non-zero weights from its internal storage list in parallel. Next, for each spatial coordinate point, the system performs a spatial weighted interpolation operation. This operation takes the extracted multiple implicit high-dimensional style prototype vector entities as input and uses the weight vector of the modulated addressing weight tensor at that coordinate point as its respective weighting coefficient. By linearly combining and summing these weighted prototype vectors, the system aggregates and generates a single, novel style transformation feature vector specific to that isolated pixel. This process is performed on a graphics processor through massively parallel computation, synthesizing a unique style vector independently for each pixel in the image. Finally, the system sequentially aggregates the style transformation feature vectors generated at all pixels along the row and column directions of the current image frame to be processed, organizing them into a tensor with the same size as the original image but with the same number of channels as the prototype vector dimension. This tensor is the final output differential style field, which constitutes the entire virtual pixel space and guides the differentiated art rendering logic at each independent point.

[0069]

[0070] The above formula is a spatial weighted interpolation operation. Wherein, In pixel coordinates The final style transformation feature vector generated at that point. The modulated addressing weights in the coordinate system The corresponding number The weight components of each prototype. The first one retrieved from the prototype library An implicit high-dimensional metastyle prototype vector entity. This represents the total number of prototype vectors. This formula smoothly transforms the discrete prototype selection process into continuous feature space interpolation by directly applying modulated weights to the prototype vectors. All vectors have a dimension of 256, and the units remain consistent throughout the calculation.

[0071] In this embodiment, the primary key for retrieval refers to a value in the modulated addressing weight vector that is greater than a very small threshold (e.g., ...). The array index or subscript of the element. Parallel decimation refers to utilizing the SIMD (Single Instruction Multiple Data) characteristic of the graphics processor to concurrently read multiple prototype vectors from the prototype library for multiple pixels, based on their different weight vectors. The style transformation feature vector is a vector with the same dimension (e.g., 256-dimensional) as the implicit high-dimensional style prototype vector. It is no longer an abstract prototype, but a set of parameters customized for a specific pixel after scene semantics, user preferences, and weighted fusion of multiple prototypes, containing specific style transformation instructions. The differential style field is the final product of this step. It is a refined style control information field that corresponds one-to-one with the original image in space. Its "differential" characteristic is reflected in the fact that the style vector at each point in the field may differ from other points, thus achieving a non-uniform stylization effect.

[0072] For example, the system processes the modulated addressing weights generated by S4. For pixel (200, 300), its modulated weight vector is [0.6, 0.2, 0.1, 0.1, 0, ...], where the first four components are non-zero. The system then uses indices 0, 1, 2, 3 as the retrieval primary key to extract the 0th, 1st, 2nd, and 3rd implicit high-dimensional style prototype vector entities in parallel from the prototype library, denoted as... Assume these prototype vectors are all 256-dimensional. Next, the system performs spatially weighted interpolation to calculate the specific style transformation feature vector for pixel (200, 300): This calculation involves vector scalar multiplication and addition performed separately along 256 dimensions. For example, if The first element is 0.5. It is -0.2. It is 0.8. The value is -0.4, then The first element is Similarly, after calculating all 256 elements, the style vector for that point is obtained. This process is performed simultaneously on all 1920x1080 pixels, ultimately generating a 1920x1080x256 differential style field, which is stored in video memory for use in the next rendering step.

[0073] S6. Read the differential style field and the current image frame to be processed, extract the multi-scale feature representation base map through a multi-layer convolutional network, use the differential style field to perform linear affine adjustment on the multi-scale feature representation base map, and inversely render the target stylized image.

[0074] In a specific embodiment of the present invention, a multi-scale feature representation base map is extracted by a multi-layer convolutional network, and a linear affine adjustment is performed on the multi-scale feature representation base map using a differential style field to invert and render the target stylized image. This includes: configuring a multi-layer convolutional network to perform an upsampling and downsampling process to transform the current image frame to be processed into a multi-scale feature representation base map containing high-dimensional, deep-level abstract feature mappings.

[0075] Style transformation feature vectors are extracted from the differential style field, and linear affine adjustment calculations are performed on the absolute mean data and standard deviation parameters of the feature channels of the multi-scale feature representation base map.

[0076] In a specific embodiment of the present invention, a linear affine adjustment calculation operation is performed on the absolute mean data and standard deviation parameters of the feature channels of the multi-scale feature representation base map, including: linearly projecting the style transformation feature vector through a fully connected layer to generate a style scaling vector and a style bias vector, respectively.

[0077] Calculate the absolute mean and standard deviation of the content feature vectors at corresponding spatial locations in the multi-scale feature representation base map, and then normalize the content feature vectors.

[0078] The normalized content feature vector is multiplied element-wise by the style scaling vector and then added element-wise by the style bias vector to complete the linear affine adjustment calculation.

[0079] The pre-built deconvolution cascade module is activated to perform upsampling and rendering on the multi-scale feature representation base map that has completed the linear affine adjustment calculation, generating the target stylized image.

[0080] Specifically, this step aims to materialize the abstract style instructions generated by S5 into the final visual image. The system first takes the current image frame with the original RGB color channels as input and feeds it into a pre-built multi-layer convolutional network. This network extracts and downsamples the image through a series of convolutional and pooling layers, transforming the high-resolution pixel data into a low-resolution multi-scale feature representation base map containing high-dimensional, deep abstract feature maps. Simultaneously, the system accesses the differential style field generated by S5 and performs corresponding spatial downsampling on the differential style field to maintain coordinate alignment, based on the spatial coordinate system of the multi-scale feature representation base map. For each spatial coordinate position in the multi-scale feature representation base map, the system extracts the values ​​of all its feature channels to form a content feature vector, and simultaneously extracts the style transformation feature vector of the differential style field at the corresponding position. Before performing the core transformation, the style transformation feature vector is first linearly projected through two independent, lightweight fully connected layers to generate a style scaling vector and a style bias vector with the same dimension as the content feature vector. Subsequently, the system performs a linear affine adjustment calculation on the content feature vector. This involves first calculating the absolute mean and standard deviation of the vector, then normalizing it using these statistics. The normalized vector is then multiplied element-wise by the style scaling vector and added element-wise by the style bias vector. This operation covers all spatial locations of the multi-scale feature representation base map, completing the injection of style information. Finally, the system activates a pre-built deconvolution cascade module with built-in pixel reconstruction decoding capabilities, using the multi-scale feature representation base map after the linear affine adjustment calculation as input. This module gradually restores the spatial resolution of the feature map to the original image size through a series of deconvolutional layers or combinations of upsampling layers and convolutional layers, while simultaneously restoring the number of feature channels to three RGB color channels, ultimately generating a target stylized image with unique scene rendering characteristics. This image is immediately written to the back buffer of the graphics hardware's double buffer and, upon the arrival of the next vertical sync signal, replaces the underlying video stream source, switching it to the foreground for display, and is immediately fed to the display panel driver interface for presentation.

[0081]

[0082] The above formula is the core equation for the linear affine adjustment calculation operation, namely adaptive instance normalization. Wherein, It is the original content feature vector at a certain spatial location in the multi-scale feature representation base map. and These are the channel-wise mean and standard deviation of the feature vector of this content, respectively. and These are the style scaling vector and style bias vector generated by projecting the style transformation eigenvector of the differential style field at that location. It is a new feature vector that has undergone affine adjustment and incorporates style information. All variables in this formula are vectors or scalars calculated from vectors, and the operations are performed in the channel dimension, maintaining consistent dimensions.

[0083] In this embodiment, the multi-layer convolutional network uses the first half of the VGG-19 network as the encoder, specifically truncated to the ReLU4_1 layer. This network has been pre-trained on a large image dataset and can effectively extract semantic content features of images. The multi-scale feature representation base map is the feature tensor output by the encoder, whose spatial size is downsampled by 8 times, while the number of channels is increased to 512. The linear affine adjustment operation is a standardization technique that aligns the distribution of content features with the distribution of style features in the feature space. It achieves style transfer by changing the mean and variance of the feature channels. The deconvolutional cascade module acts as a decoder, whose structure is topologically symmetrical with the encoder. It is used to reconstruct an image with realistic texture and color from the stylized feature map. The target stylized image is the final output of the entire processing flow. It is an RGB image with the same size as the original input image. Its content structure originates from the original image, but the color and texture style have been reconstructed by the differential style location.

[0084] For example, the system receives a 1920x1080 pixel current image frame to be processed and inputs it into the VGG-19 encoder to obtain a 240x135x512 multi-scale feature representation base map. Simultaneously, the system reads a 1920x1080x256 differential style field generated by S5. Consider the 512-dimensional content feature vector at position (30,50) on the feature base map. The system calculates its mean. The standard deviation is 12.5. The system is rated at 8.0. It also extracts a 256-dimensional style transformation feature vector from the differential style field at positions (30*8, 50*8), i.e., (240, 400). This is then fed into two fully connected layers to obtain a 512-dimensional style scaling vector. and style bias vector Assuming a vector The first channel value is 20.5, vector The first channel value is 1.5, vector The first channel value is -4.0. Therefore, after the linear affine adjustment calculation, the new feature vector... The first channel value will become This computation is performed in parallel across all 512 channels and all 240x135 spatial locations. The adjusted 240x135x512 feature map is fed into the decoder, which performs three upsampling and convolution operations to finally render a 1920x1080x3 target stylized image. This image is immediately transmitted to the back buffer of the display system and displayed on the screen when the next frame is refreshed, completing the real-time stylization replacement of the original video stream.

[0085] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, and all such modifications and additions should fall within the protection scope of the present invention.

Claims

1. A cross-scene special effects style adaptive method based on meta-learning, characterized in that, Includes the following steps: S1. Obtain style feature asset data, build a prototype library containing implicit high-dimensional style prototype vectors in memory, load the pre-trained style field generation network, establish a communication link between the prototype library and the style field generation network, and generate a style adaptive system. S2. Receive the current image frame to be processed, parse the pixel structure and dependency relationship of the current image frame to be processed, extract multi-dimensional scene perception data, synthesize and output a customized structured semantic descriptor; S3. Import the customized structured semantic descriptor into the style adaptation system, perform forward mapping and similarity calculation through the style field generation network, and generate addressing weights pointing to the implicit high-dimensional style prototype vector. S4. Obtain the human subjective style intensity configuration parameters, and superimpose the human subjective style intensity configuration parameters with the addressing weights to generate the modulated addressing weights. S5. Based on the modulated addressing weights, extract the corresponding implicit high-dimensional style prototype vectors from the prototype library, and synthesize the differential style field for all pixels of the current image frame through spatial weighted interpolation. S6. Read the differential style field and the current image frame to be processed, extract the multi-scale feature representation base map through a multi-layer convolutional network, use the differential style field to perform linear affine adjustment on the multi-scale feature representation base map, and inversely render the target stylized image.

2. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The establishment of a communication link between the prototype library and the style field generation network to generate a style adaptive system includes: A prototype library containing multiple sets of implicit high-dimensional style prototype vectors is constructed in a solid-state storage space. The implicit high-dimensional style prototype vectors represent potential universal style transformation preference parameters that are independent of a single basic art filter framework. The style field generation network, which has been pre-trained offline, is loaded into the central processing unit's running memory. The style field generation network serves as a data processing engine for establishing feature addressing and solving mechanisms for corresponding prototype library resources. A data reading and instruction exchange communication link is established between the prototype library and the style field generation network on the underlying hardware motherboard bus to generate a style adaptive system in a memory-resident, ready-to-go state.

3. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The process of parsing the pixel structure and dependencies of the current image frame to be processed, extracting multi-dimensional scene perception data, and synthesizing and outputting a customized structured semantic descriptor includes: Activate the edge detection convolution operator to detect the pixel transition smoothing level within the current image frame to be processed, and generate a semantic gradient map in spatial dimension; The saliency detection algorithm is run to evaluate the content focus of the current image frame to be processed, segment and define the preserved region and the transformed region, and output the content sensitivity mask; Measure the numerical distances of semantic objects in the two-dimensional coordinate plane and match their spatial binding tightness to construct a region association graph; Based on the feature depth dimension, the semantic gradient map, content sensitivity mask and region association map are connected to generate a customized structured semantic descriptor.

4. The cross-scene special effects style adaptation method based on meta-learning according to claim 3, characterized in that, The method measures the numerical values ​​of the arrangement distances of semantic objects on the two-dimensional coordinate plane and matches their spatial binding tightness to construct a region association map, including: The Euclidean distance equation is used to measure the numerical values ​​of the arrangement distances of semantic objects with different structural attributes in a two-dimensional coordinate plane. Retrieve a pre-defined table of co-occurrence probabilities for each category and match the spatial binding density of each region; The numerical values ​​of the arrangement distance are fused with the spatial density and encoded to construct a regional association map.

5. The cross-scene special effects style adaptation method based on meta-learning according to claim 3, characterized in that, The output content sensitivity mask includes: The central processing unit is controlled to run a saliency detection algorithm to calculate the content focus score for each pixel position in the current image frame to be processed; The content focus score is compared pixel by pixel with the preset safety tolerance score. Pixels with content focus scores higher than the safety tolerance score are assigned to the reserved area, while pixels with content focus scores lower than the safety tolerance score are assigned to the transform area, thereby generating a binarized content sensitivity mask.

6. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The step of generating addressing weights pointing to implicit high-dimensional style prototype vectors through forward mapping and similarity calculation via a style field generation network includes: In the pre-perception layer of the style field generation network, a fully connected forward mapping propagation is performed on the customized structured semantic descriptor to extract the spatial semantic tensor; The set of cosine similarity values ​​between the spatial semantic tensor and the corresponding implicit high-dimensional style prototype vectors is calculated by vector multiplication. A nonlinear normalization function is introduced to process the set of cosine similarity values ​​and compress it into the standard probability space, generating addressing weights that point to implicit high-dimensional style prototype vectors.

7. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The step of obtaining the human subjective style intensity configuration parameter, superimposing the human subjective style intensity configuration parameter with the addressing weight, and generating the modulated addressing weight includes: A low-level operating system interrupt response listener is built to capture intensity change update events triggered by graphical touch interaction components, and to extract human subjective style intensity configuration parameters. Analyze the pixel content sensitivity of the current image frame to be processed, and calculate the region scaling compensation factor with anti-boundary distortion properties; The executor multiplies the subjective style intensity configuration parameter with the regional scaling compensation factor to obtain the modulation scalar; The modulation scalar is mathematically superimposed into the addressing weight matrix dimension, and the result is normalized in the feature channel dimension to ensure that the sum of the weights is 1, thus generating the modulated addressing weights.

8. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The step of synthesizing a differential style field for all pixels of the current image frame through spatial weighted interpolation includes: Using the non-zero array index of the modulated addressing weights as the retrieval key, the corresponding implicit high-dimensional style prototype vectors are extracted in parallel from the prototype library. The matrix eigenvalues ​​of the implicit high-dimensional style prototype vector are extracted and combined with the modulated addressing weights to perform spatial weighted interpolation operations, and aggregated to generate style transformation feature vectors for individual isolated pixels. The style transformation feature vector clusters generated globally are progressively advanced along the row and column directions of the current image frame to be processed and converged to form a differential style field.

9. The cross-scene special effects style adaptation method based on meta-learning according to claim 1, characterized in that, The process involves extracting a multi-scale feature representation base map using a multi-layer convolutional network, applying a differential style field to the multi-scale feature representation base map via linear affine adjustment, and then inverting and rendering the target stylized image, including: Configure a multi-layer convolutional network to perform an upsampling and downsampling process, transforming the current image frame to be processed into a multi-scale feature representation base map containing high-dimensional, deep-level abstract feature maps; Extract the style transformation feature vector within the differential style field, and perform linear affine adjustment calculations on the absolute mean data and standard deviation parameters of the feature channels of the multi-scale feature representation base map. The pre-built deconvolution cascade module is activated to perform upsampling and rendering on the multi-scale feature representation base map that has completed the linear affine adjustment calculation, generating the target stylized image.

10. The cross-scene special effects style adaptation method based on meta-learning according to claim 9, characterized in that, The linear affine adjustment calculation operation for the absolute mean data and standard deviation parameter of the feature channels of the multi-scale feature representation base map includes: The style transformation feature vector is linearly projected through a fully connected layer to generate style scaling vector and style bias vector, respectively. Calculate the absolute mean and standard deviation of the content feature vectors of the multi-scale feature representation base map at the corresponding spatial location, and normalize the content feature vectors. The normalized content feature vector is multiplied element-wise by the style scaling vector and then added element-wise by the style bias vector to complete the linear affine adjustment calculation.

Images