A rapid method for generating customizable cultural IP derivatives

Abstract

This invention belongs to the field of cultural IP derivative product generation technology, specifically a rapid generation method for customizable cultural IP derivative products. Addressing the limitations of existing methods that fail to achieve end-to-end automated generation from voice commands to printable derivative products, lack semantic coordination and style consistency across multiple IP characters and scenes, and cannot support independent interaction and joint editing of multiple characters and accessories within multiple display areas, the following solution is proposed, comprising the following steps: S1: Voice command recognition and target information generation. This invention enables end-to-end automated generation from voice commands to cultural IP derivative product images; it can simultaneously process multiple IP characters and target scenes, achieving semantic coordination and style consistency; and it supports real-time personalized adjustments to characters, accessories, and colors by users, ultimately generating high-quality IP cultural derivative product images with consistent style and harmonious lighting.

Description

Technical Field

[0001] This invention relates to the field of cultural IP derivative product generation technology, and in particular to a rapid method for generating customizable cultural IP derivative products. Background Technology

[0002] With the rapid development of the cultural and creative industries and the virtual IP ecosystem, users' demands for personalized and interactive cultural derivatives are constantly increasing. Currently, the generation of IP derivatives often relies on manual design and the splicing of single materials, resulting in long production cycles, weak interactivity, low automation, and difficulty in supporting the personalized combination of multiple characters, dynamic scenes, and accessories. Although existing speech recognition and AI image generation technologies can achieve image synthesis to some extent, they still have the following shortcomings in semantic-driven, multi-character collaboration, and real-time customization:

[0003] 1. It cannot achieve end-to-end automated generation from voice commands to printable derivatives; 2. It lacks semantic coordination and style consistency processing between multiple IP characters and scenes; 3. The existing system cannot support users to independently interact with and jointly edit multiple characters and accessories in multiple display areas;

[0004] In summary, this application proposes a method for the rapid generation of customizable cultural IP derivatives. Summary of the Invention

[0005] This invention proposes a rapid generation method for customizable cultural IP derivatives, which solves the problems in existing technologies such as the inability to achieve end-to-end automated generation from voice commands to printable derivatives, the lack of semantic coordination and style consistency processing between multiple IP characters and scenes, and the inability of existing systems to support users to independently interact with and jointly edit multiple characters and accessories in multiple display areas.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A rapid method for generating customizable cultural IP derivatives includes the following steps:

[0008] S1: Voice command recognition and target information generation: The terminal device uses Tencent Cloud's recognition engine and natural speech processing (NLP) technology to recognize and extract the user's voice commands, generate corresponding target information, and send it to the local server.

[0009] S2: Semantic parsing and multi-role scene retrieval: The local server uses the BERT model to perform semantic parsing on the target information in S1, and obtains the parsing results. The parsing results include at least one IP role identifier, target scene keywords and generation instruction type. Based on the parsing results, the local server performs semantic vector retrieval in the locally stored IP role database and scene database using the Faiss approximate nearest neighbor retrieval algorithm. Cosine similarity and weighted fusion similarity are used as matching metrics. At the same time, the image style feature consistency index is combined for comprehensive evaluation to obtain several target role image sets that match the target information and target scene images that are consistent with the target information. The target scene images are then sent to the terminal device.

[0010] S3: Send a request to retrieve associated component resources: The local server sends a request to the cloud server to retrieve associated components based on the IP role identifiers parsed in S2.

[0011] S4: Response to Request: The cloud server responds to the request by retrieving the accessory set resource package corresponding to each IP role identifier via API call and sending it to the terminal device;

[0012] S5: Multi-window display and interactive activation: The terminal device displays the target scene image in the main display box, several target character images in the first sub-display box, and the corresponding accessory set in the second sub-display box. Users can activate any character or accessory in the sub-display box through touch operation. After activation, the terminal system embedded in the terminal device automatically establishes the association channel between the character and the accessory in the main display box. After establishment, the selected character image is placed on the surface of the target scene image. The position and proportion of the character are adjusted by gesture in the target scene image. Users can select the corresponding clothing, decoration or color change item in the accessory set by touch and apply it to the corresponding character image in the main display box in real time.

[0013] S6: Send a compositing request: After the user completes all role and scene configurations, the "composite" button on the terminal device is triggered, and the terminal device sends a multi-image compositing request to the local server;

[0014] S7: Image Compositing Processing: The local server uses multi-resolution fusion technology, Poisson fusion algorithm and gain compensation algorithm to fuse the images of each character, accessory and scene, eliminate seams and color difference, generate IP cultural derivative images with unified style and lighting, and send them to the terminal device.

[0015] S8: Output and Printing: The terminal device displays the IP cultural derivative images generated in S7 in the main display box, and controls the printing device to print the physical output via wireless communication.

[0016] Preferably, the specific logical steps of S1 are as follows:

[0017] S101: The terminal device acquires the user's voice signal s(t) in real time through the built-in microphone array, and performs noise suppression processing on the acquired raw voice signal to remove silent segments and environmental noise;

[0018] The formula used for noise suppression using spectral subtraction is:

[0019] ;

[0020] in This is the speech spectrum after noise reduction. The spectrum of the original speech signal is represented in the frequency domain by performing a Fast Fourier Transform on s(t), reflecting the amplitude and phase of each frequency component. The spectrum of the input speech. For noise estimation, This is the noise compensation coefficient;

[0021] S102: Analyze the spectrum of the denoised speech. The process involves frame segmentation, windowing, and Fast Fourier Transform, followed by extraction of acoustic feature vectors. ,in All are vector elements, representing the acoustic feature vector of the speech in the t-th frame;

[0022] S103: Input the acoustic feature vector X into the Tencent Cloud speech recognition model and calculate the acoustic probability. With language probability The optimal word sequence is obtained through Viterbi decoding. The output is the transcribed text T;

[0023] The formula used above is as follows:

[0024] ;

[0025] Where X = { } represents the acoustic eigenvector. Let be the acoustic feature vector of the t-th frame. When in a hidden state, it represents a phoneme or subphoneme. For the probability of launch, Ta is the state transition probability, and Ta is the number of speech frames, i.e., the total number of frames after the speech signal is divided into frames.

[0026] ;

[0027] Where W = { } represents the word sequence corresponding to the speech recognition result. Let the order of the language model be denoted as . For the i-th word;

[0028] Where P(W|X) = , This indicates the search for the word sequence W that maximizes the joint probability. The optimal word sequence is the text sequence that the terminal system ultimately outputs.

[0029] T=ASR ( )={ }, where ASR ( ) represents the automatic speech recognition function, X represents the acoustic feature sequence of the speech signal, and T represents the target text output by the speech recognition, i.e., the transcribed text. } represents the word sequence corresponding to the speech recognition result;

[0030] S104: Input the transcribed text T from S103 into the Tencent Cloud NLP intent recognition model for word segmentation, part-of-speech tagging, and semantic vectorization. The formula used for semantic vectorization is:

[0031] ;

[0032] in Let i be the word vector of the i-th word. It is a semantic vector;

[0033] Tencent Cloud NLP intent recognition model outputs intent categories based on classification probabilities, using the following formula: ;

[0034] Where I represents the set of candidate intents. Indicates the category of intent;

[0035] S105: When Intent Category I * Once determined, the terminal system extracts key slots from the semantic dependency tree. These key slots include objects, operations, and attributes, generating structured target information. ;

[0036] in For IP role identification, Keywords for the target scenario To generate instruction types, For semantic vectors, Context-related parameters;

[0037] S106: The terminal encapsulates the generated target information G into a JSON format message and sends it to the local server through a secure channel.

[0038] Preferably, the specific logical steps of S2 are as follows:

[0039] S201: The local server receives target information G from the terminal device. The text or key fields in G are then input into the BERT model for deep semantic parsing, and the parsing results are obtained:

[0040] PaseResult={ , , };

[0041] ;

[0042] in ( () represents the semantic encoding function based on the BERT model, used to map input text into high-dimensional semantic vectors. The input text is a description of a character or scene. This represents the text description information corresponding to the i-th IP role identifier. This represents the semantic feature vector of the i-th IP role text after BERT encoding. Represents the set of keywords for the target scenario. This represents the semantic feature vector obtained by encoding scene keywords using the BERT model;

[0043] S202: Store the semantic vector of each image in the locally stored IP role library and scene library respectively: RoleDB=( SceneDB=( RoleDB represents the local IP character library, storing all character images and their corresponding semantic vectors; SceneDB represents the local scene image library, storing all scene images and their semantic vectors. This represents the j-th character image file. This indicates the relationship with the j-th character image. The corresponding semantic vector, This represents the k-th scene image file. Represents the scene image with respect to the k-th image. The corresponding semantic vector;

[0044] Use the Faiss approximate nearest neighbor search algorithm to find the most similar vector: , ;

[0045] in Represents the semantic vector of the i-th IP role. The set of most similar character image indexes. Representation and scene semantic vector The index of the most similar scene images, ( The similarity calculation function can be cosine similarity, the reciprocal of Euclidean distance, or a weighted fusion similarity function. This indicates that the index of the character sample with the highest similarity is selected from the database. This indicates that the scene sample index with the highest similarity is selected from the scene library;

[0046] S203: Similarity is calculated using cosine similarity and weighted fusion similarity, with the following formula: Cosine similarity: cos_sim(A,B) =

[0047] Where A represents the semantic vector of the query target, and B represents the semantic vector of the candidate images in the database.

[0048] Represents the vector product point. The vector is represented by the Euclidean norm, and cos_sim(A,B) represents the cosine similarity between two vectors in the semantic space, with a value range of [-1,1].

[0049] Weighted fusion similarity: sim_total= ,in and These are the weighting coefficients. To calculate the image style consistency index, the image style consistency index includes color, lighting and texture. sim_total represents the comprehensive similarity score, which is used for sorting and filtering candidate images.

[0050] S204: Sort the candidate image set using the following formula:

[0051] ,in This is a sorted collection of character images;

[0052] The top H character images that are consistent with the target information and have a high style match are retained, using the following formula:

[0053] SelectedRoles = top_H(Ranked_Roles), where H is the upper limit of the number of candidates to be retained, and top_H ( ) represents the function that selects the top H elements, and SelectedRoles represents the set of H character images with the highest similarity;

[0054] A similar filtering process is performed on the collection of scene images, using the following formula:

[0055] Where SelectedScene represents the target scene image with the highest similarity. This indicates that the parameter K is chosen to maximize the objective function. It represents the comprehensive similarity between the target semantics and the scene image;

[0056] S205: Send the final selected target scene image to the main display frame of the terminal device, and send the selected target character image set to the first secondary display frame of the terminal device.

[0057] Preferably, in step S3, when sending the associated component, it is necessary to send each IP Role Identifier Construction Component Retrieval Request Packet:

[0058] ;

[0059] in This serves as an identifier for the current task scenario. This serves as a model feature fingerprint, used to identify the correspondence between local matching results and cloud resources. To request a security token, used to verify identity and prevent unauthorized calls;

[0060] Before sending, the local server needs to encrypt the retrieval request packet using the HTTPS encrypted messaging protocol. The encryption method uses the following formula:

[0061] Send( )= ;

[0062] in This indicates that the SM4 symmetric encryption algorithm, a national standard, is used. The session key is Send(), which is a request to transmit data.

[0063] Preferably, the specific logical steps of S4 are as follows:

[0064] S401: The cloud server received a related component retrieval request packet from the local server. Then, perform the decryption and verification operations: And conduct a legality check.

[0065] Verify )= ;

[0066] in This is the SM4 symmetric decryption function. `Verify()` is the verification module used to check the request signature and permissions. To request a security token, For session key;

[0067] S402: When the verification result is true, the cloud server determines the IP role identifier. Determine the corresponding API interface and resource path in the resource mapping table: ,in This is a function that maps character identifiers to resource paths. The interface endpoint being invoked;

[0068] S403: The cloud server retrieves the IP role identifier by calling the corresponding database or object storage via API. Corresponding accessory set resource pack:

[0069] ;

[0070] in A represents the original set of parts matched in the cloud, and A is the parts resource library. These are the semantic feature vectors of the IP character and the accessory set, respectively, and sim() is the similarity calculation function: sim( , )= E ( ) is a vector encoding function. For the semantic feature vector of the IP role, This represents the semantic feature vector of the candidate accessory set;

[0071] S404: Cloud server retrieves the set of accessories. Tagging and encapsulation are performed to build resource packages with a unified format: ,in For packaged structured component sets, This is the first color change set used for changing clothing colors. Replacement set for decorations A set for changing the colors of decorative items. This is the model metadata, which includes component dimensions, hierarchical relationships, and compatibility information;

[0072] S405: The cloud server encrypts the structured resource packet and sends it back to the terminal device via an API response, Resp( )= ( , Once the terminal device receives the data, it displays the accessory set in the second display box for user interaction.

[0073] Preferably, the specific logical steps of S5 are as follows:

[0074] S501: The terminal device receives the target scene image from the local server. After the target character image set R and accessory set Perform the following operations: Load the target scene image in the main display box. Load the target character image set R= in the first display box. ,in All are set elements, representing the nth character image in the target character image set. The corresponding accessory set is loaded in the second display box. ={ },in All are elements of a set, representing the nth accessory material in the accessory set;

[0075] S502: When a user interacts with a target character or accessory in the sub-display box via touch or gesture, the terminal system captures the touch event. And determine the activation state: Active(x) = ;in To activate the threshold, Active(x)=1 indicates the activated object. After activation, the terminal system automatically establishes a role-accessory association channel in the main display box.

[0076] S503: When a target character image is detected When activated, the terminal system establishes an association mapping: ,in For character-accessory mapping channel, To match the target character image The corresponding accessory set; this mapping channel is used to synchronize the character's accessory replacement logic in the main display frame.

[0077] S504: The terminal system will activate the target character image. Drag and drop from the first secondary display frame into the main display frame, and overlay it onto the target scene image. Surface layer: ,in For the current combined layer, This indicates an overlay operation. Users can adjust the position and proportion of the character using gestures. The terminal system updates the character's state based on the gesture trajectory parameters (Δx, Δy, λ), using the following formula: ,in For gesture movement offset, For scaling ratio, For translation and scaling transformation functions;

[0078] S505: When the user selects an accessory item in the second sub-display box ∈ At that time, the terminal system applies it to the target character image in real time. The above steps involve style blending and color adjustments during rendering, using the following formula: ,in For color adjustment functions, For texture feature extraction, and For fusion weights;

[0079] S506: After the user completes the character accessory combination, the terminal system stores the current interaction state in a structured format. ,in This is the current combination state. This is the current interaction mode. For timestamps;

[0080] S507: The terminal system supports multiple characters being activated simultaneously and bound to their respective accessory sets, forming a multi-character, multi-channel mapping: Users can adjust the appearance and position of multiple characters in parallel, enabling dynamic generation of derivative images where multiple IP characters coexist in the same scene.

[0081] Preferably, the specific logical steps of S7 are as follows:

[0082] S701: The local server receives the resource package of character images, accessory images and scene background images sent by the cloud server, and performs preprocessing operations on each image, including size standardization, format unification, background transparency and feature point extraction.

[0083] S702: The key point mapping relationship between the character and accessory images is calculated using SIFT or ORB feature matching algorithms, and the position alignment and scale correction are completed by using the affine transformation matrix T(x,y). The formula used is:

[0084] ;

[0085] in This represents the affine transformation matrix, which includes translation, rotation, and scaling parameters;

[0086] S703: Multi-resolution image decomposition is performed using the Laplacian pyramid and Gaussian pyramid, and weighted fusion is applied to each layer of the image. The formula used is as follows:

[0087] ;

[0088] in This is the image after fusion at layer l. and Laplace layers for the characters and background, respectively. and For each layer, a blending weight mask is used to effectively smooth the stitching edges and maintain detail consistency through layered blending; S704: Poisson blending is performed on the blending region, and a continuous transition between color and brightness is achieved by solving the Poisson equation. The formula used is:

[0089] ;

[0090] Where g is the source image Ω represents the target region image, and Ω represents the fusion region.

[0091] S705: Calculate the luminance and chromaticity statistics of the fused image and adjust the gain for luminance differences. ;

[0092] in This is the gain coefficient. As offset compensation parameters, histogram equalization and gamma correction are used to achieve overall style consistency and brightness balance;

[0093] S706: After fusion, an IP cultural derivative image with a unified style and consistent lighting is generated. The local server compresses and encodes the resulting image and sends it to the terminal device through the local transmission protocol.

[0094] Compared with the prior art, the beneficial effects of the present invention are:

[0095] 1. In this invention, through speech recognition and semantic parsing technology, semantic intent can be automatically identified from user voice commands, generating structured target information. Through the BERT semantic parsing model and Faiss semantic retrieval algorithm, end-to-end fully automated processing from voice input to IP derivative image output is achieved, significantly improving generation efficiency and reducing manual intervention.

[0096] 2. In this invention, through semantic vector matching, cosine similarity and weighted fusion similarity comprehensive evaluation mechanism, multiple character identifiers and target scene keywords can be parsed simultaneously to achieve semantic coordination matching of multiple characters and multiple scenes, ensuring consistency and semantic coherence between characters and scene styles;

[0097] 3. In this invention, a main display frame and two secondary display frames are set on the terminal device to display the scene, character and accessory set respectively. Users can activate the character, replace the accessory and customize the color through touch and gesture operation, and adjust the position and proportion in the main display frame in real time, thereby realizing a highly free interactive customization experience.

[0098] 4. In this invention, by employing multi-resolution fusion, Poisson fusion and gain compensation algorithms, the characters, accessories and scenes are seamlessly synthesized and the lighting and color are uniformly corrected, effectively eliminating seams and color differences, and ensuring that the final generated cultural IP derivative images have visual style consistency and high-quality expressiveness.

[0099] This invention combines speech recognition, natural language processing, semantic retrieval, and multimodal image fusion technology to achieve end-to-end automated generation of images from voice commands to cultural IP derivative products. This overcomes the problems of existing technologies, such as reliance on manual design, long generation cycles, and weak interactivity. It can simultaneously process multiple IP characters and target scenes, achieving semantic coordination and style consistency. Furthermore, through a multi-window interactive interface, it supports users to make real-time personalized adjustments to characters, accessories, and colors. Finally, the fusion algorithm generates high-quality IP cultural derivative product images with consistent style and coordinated lighting, thereby significantly improving the generation efficiency, intelligence level, and interactive experience of cultural derivative products. Attached Figure Description

[0100] Figure 1 This is a sequence diagram of a rapid generation method for customizable cultural IP derivatives proposed in this invention.

[0101] Figure 2 This is a diagram of the interface of the terminal device in the method for rapidly generating customizable cultural IP derivatives proposed in this invention. Detailed Implementation

[0102] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0103] Reference Figure 1-2 This embodiment proposes a rapid method for generating customizable cultural IP derivatives, including the following steps:

[0104] S1: Voice command recognition and target information generation: The terminal device uses Tencent Cloud's recognition engine and natural speech processing (NLP) technology to recognize and extract the user's voice commands, generate corresponding target information, and send it to the local server.

[0105] S2: Semantic parsing and multi-role scene retrieval: The local server uses the BERT model to perform semantic parsing on the target information in S1, and obtains the parsing results. The parsing results include at least one IP role identifier, target scene keywords and generation instruction type. Based on the parsing results, the local server performs semantic vector retrieval in the locally stored IP role database and scene database using the Faiss approximate nearest neighbor retrieval algorithm. Cosine similarity and weighted fusion similarity are used as matching metrics. At the same time, the image style feature consistency index is combined for comprehensive evaluation to obtain several target role image sets that match the target information and target scene images that are consistent with the target information. The target scene images are then sent to the terminal device.

[0106] S3: Send a request to retrieve associated component resources: The local server sends a request to the cloud server to retrieve associated components based on the IP role identifiers parsed in S2.

[0107] When sending associated components, it is necessary to do so for each one. IP Role Identifier Construction Component Retrieval Request Packet:

[0108] ;

[0109] in This serves as an identifier for the current task scenario. This serves as a model feature fingerprint, used to identify the correspondence between local matching results and cloud resources. To request a security token, used to verify identity and prevent unauthorized calls;

[0110] Before sending, the local server needs to encrypt the retrieval request packet using the HTTPS encrypted messaging protocol. The encryption method uses the following formula:

[0111] Send( )= ;

[0112] in This indicates that the SM4 symmetric encryption algorithm, a national standard, is used. The session key is `Send()`, which is a request to transmit data.

[0113] S4: Response to Request: The cloud server responds to the request by retrieving the accessory set resource package corresponding to each IP role identifier via API call and sending it to the terminal device;

[0114] S5: Multi-window display and interactive activation: The terminal device displays the target scene image in the main display box, several target character images in the first sub-display box, and the corresponding accessory set in the second sub-display box. Users can activate any character or accessory in the sub-display box through touch operation. After activation, the terminal system embedded in the terminal device automatically establishes the association channel between the character and the accessory in the main display box. After establishment, the selected character image is placed on the surface of the target scene image. The position and proportion of the character are adjusted by gesture in the target scene image. Users can select the corresponding clothing, decoration or color change item in the accessory set by touch and apply it to the corresponding character image in the main display box in real time.

[0115] S6: Send a compositing request: After the user completes all role and scene configurations, the "composite" button on the terminal device is triggered, and the terminal device sends a multi-image compositing request to the local server;

[0116] S7: Image Compositing Processing: The local server uses multi-resolution fusion technology, Poisson fusion algorithm and gain compensation algorithm to fuse the images of each character, accessory and scene, eliminate seams and color difference, generate IP cultural derivative images with unified style and lighting, and send them to the terminal device.

[0117] S8: Output and Printing: The terminal device displays the IP cultural derivative images generated in S7 in the main display box, and controls the printing device to print physical outputs via wireless communication.

[0118] The printed products are 3D stereoscopic products.

[0119] In this implementation scheme, the specific logical steps of step S1 are as follows:

[0120] S101: The terminal device acquires the user's voice signal s(t) in real time through the built-in microphone array, and performs noise suppression processing on the acquired raw voice signal to remove silent segments and environmental noise;

[0121] The formula used for noise suppression using spectral subtraction is:

[0122] ;

[0123] in This is the speech spectrum after noise reduction. The spectrum of the original speech signal is represented in the frequency domain by performing a Fast Fourier Transform on s(t), reflecting the amplitude and phase of each frequency component. The spectrum of the input speech. For noise estimation, This is the noise compensation coefficient;

[0124] S102: Analyze the spectrum of the denoised speech. The process involves frame segmentation, windowing, and Fast Fourier Transform, followed by extraction of acoustic feature vectors. ,in All are vector elements, representing the acoustic feature vector of the speech in the t-th frame;

[0125] S103: Input the acoustic feature vector X into the Tencent Cloud speech recognition model and calculate the acoustic probability. With language probability The optimal word sequence is obtained through Viterbi decoding. The output is the transcribed text T;

[0126] ;

[0127] Where X = {

[0128] } represents the acoustic eigenvector.

[0129] Let be the acoustic feature vector of the t-th frame.

[0130] When in a hidden state, it represents a phoneme or subphoneme.

[0131] For the probability of launch,

[0132] Ta is the state transition probability, and Ta is the number of speech frames, i.e., the total number of frames after the speech signal is divided into frames.

[0133] ;

[0134] Where W = { } represents the word sequence corresponding to the speech recognition result. Let the order of the language model be denoted as . For the i-th word;

[0135] Where P(W|X) = , This indicates the search for the word sequence W that maximizes the joint probability. The optimal word sequence is the text sequence that the terminal system ultimately outputs.

[0136] T=ASR ( )={ }, where ASR ( ) represents the automatic speech recognition function, X represents the acoustic feature sequence of the speech signal, and T represents the target text output by the speech recognition, i.e., the transcribed text. } represents the word sequence corresponding to the speech recognition result;

[0137] S104: Input the transcribed text T from S103 into the Tencent Cloud NLP intent recognition model for word segmentation, part-of-speech tagging, and semantic vectorization. The formula used for semantic vectorization is:

[0138] ;

[0139] in Let i be the word vector of the i-th word. It is a semantic vector;

[0140] Tencent Cloud NLP intent recognition model outputs intent categories based on classification probabilities, using the following formula: ;

[0141] Where I represents the set of candidate intents. Indicates the category of intent;

[0142] S105: When Intent Category I *Once determined, the terminal system extracts key slots from the semantic dependency tree. These key slots include objects, operations, and attributes, generating structured target information. ;

[0143] in

[0144] For IP role identification,

[0145] Keywords for the target scenario

[0146] To generate instruction types, For semantic vectors, Context-related parameters;

[0147] S106: The terminal encapsulates the generated target information G into a JSON format message and sends it to the local server through a secure channel.

[0148] In this implementation scheme, the specific logical steps of step S2 are as follows:

[0149] S201: The local server receives target information G from the terminal device. The text or key fields in G are then input into the BERT model for deep semantic parsing, and the parsing results are obtained:

[0150] PaseResult={ , , };

[0151] ;

[0152] in ( () represents the semantic encoding function based on the BERT model, used to map input text into high-dimensional semantic vectors. The input text is a description of a character or scene. This represents the text description information corresponding to the i-th IP role identifier. This represents the semantic feature vector of the i-th IP role text after BERT encoding. Represents the set of keywords for the target scenario. This represents the semantic feature vector obtained by encoding scene keywords using the BERT model;

[0153] S202: Store the semantic vector of each image in the locally stored IP role library and scene library respectively: RoleDB=( SceneDB=( RoleDB represents the local IP character library, storing all character images and their corresponding semantic vectors; SceneDB represents the local scene image library, storing all scene images and their semantic vectors. This represents the j-th character image file. This indicates the relationship with the j-th character image. The corresponding semantic vector, This represents the k-th scene image file. Represents the scene image with respect to the k-th image. The corresponding semantic vector;

[0154] Use the Faiss approximate nearest neighbor search algorithm to find the most similar vector: , ;

[0155] in Represents the semantic vector of the i-th IP role. The set of most similar character image indexes. Representation and scene semantic vector The index of the most similar scene images, ( The similarity calculation function can be cosine similarity, the reciprocal of Euclidean distance, or a weighted fusion similarity function. This indicates that the index of the character sample with the highest similarity is selected from the database. This indicates that the scene sample index with the highest similarity is selected from the scene library;

[0156] S203: Similarity is calculated using cosine similarity and weighted fusion similarity, with the following formula: Cosine similarity: cos_sim(A,B) = Where A represents the semantic vector of the query target, and B represents the semantic vector of the candidate images in the database. Represents the vector product point. The vector is represented by the Euclidean norm, and cos_sim(A,B) represents the cosine similarity between two vectors in the semantic space, with a value range of [-1,1].

[0157] Weighted fusion similarity: ,in and These are the weighting coefficients. To calculate the image style consistency index, the image style consistency index includes color, lighting and texture. sim_total represents the comprehensive similarity score, which is used for sorting and filtering candidate images.

[0158] S204: Sort the candidate image set using the following formula:

[0159] ,in This is a sorted collection of character images;

[0160] The top H character images that are consistent with the target information and have a high style match are retained. The formula used is: SelectedRoles = top_H(Ranked_Roles), where H is the upper limit of the number of candidates to be retained, and top_H ( ) represents the function that selects the top H elements, and SelectedRoles represents the set of H character images with the highest similarity;

[0161] A similar filtering process is performed on the collection of scene images, using the following formula:

[0162] Where SelectedScene represents the target scene image with the highest similarity. This indicates that the parameter K is chosen to maximize the objective function. It represents the comprehensive similarity between the target semantics and the scene image;

[0163] S205: Send the final selected target scene image to the main display frame of the terminal device, and send the selected target character image set to the first secondary display frame of the terminal device.

[0164] In this implementation scheme, the specific logical steps of step S4 are as follows:

[0165] S401: The cloud server received a related component retrieval request packet from the local server. Then, perform the decryption and verification operations: And conduct a legality check.

[0166] Verify )= ;

[0167] in This is the SM4 symmetric decryption function. `Verify()` is the verification module used to check the request signature and permissions. To request a security token, For session key;

[0168] S402: When the verification result is true, the cloud server determines the IP role identifier. Determine the corresponding API interface and resource path in the resource mapping table: ,in This is a function that maps character identifiers to resource paths. The interface endpoint being invoked;

[0169] S403: The cloud server retrieves the IP role identifier by calling the corresponding database or object storage via API. Corresponding accessory set resource pack:

[0170] ;

[0171] in A represents the original set of parts matched in the cloud, and A is the parts resource library. These are the semantic feature vectors of the IP character and the accessory set, respectively, and sim() is the similarity calculation function: sim( , )= E ( ) is a vector encoding function. For the semantic feature vector of the IP role, This represents the semantic feature vector of the candidate accessory set;

[0172] S404: Cloud server retrieves the set of accessories. Tagging and encapsulation are performed to build resource packages with a unified format: ,in For packaged structured component sets, This is the first color change set used for changing clothing colors. Replacement set for decorations A set for changing the colors of decorative items. This is the model metadata, which includes component dimensions, hierarchical relationships, and compatibility information;

[0173] S405: The cloud server encrypts the structured resource packet and sends it back to the terminal device via an API response, Resp( )= ( , Once the terminal device receives the data, it displays the accessory set in the second display box for user interaction.

[0174] In this implementation scheme, the specific logical steps of step S5 are as follows:

[0175] S501: The terminal device receives the target scene image from the local server. After the target character image set R and accessory set Perform the following operations: Load the target scene image in the main display box. Load the target character image set R= in the first display box. ,in All are set elements, representing the nth character image in the target character image set, and the corresponding accessory set is loaded in the second display box. ={ },in All are elements of a set, representing the nth accessory material in the accessory set;

[0176] S502: When a user interacts with a target character or accessory in the sub-display box via touch or gesture, the terminal system captures the touch event. And determine the activation state: Active(x) = ;in To activate the threshold, Active(x)=1 indicates the activated object. After activation, the terminal system automatically establishes a role-accessory association channel in the main display box.

[0177] S503: When a target character image is detected When activated, the terminal system establishes an association mapping: ,in For character-accessory mapping channel, To match the target character image The corresponding accessory set; this mapping channel is used to synchronize the character's accessory replacement logic in the main display frame.

[0178] S504: The terminal system will activate the target character image. Drag and drop from the first secondary display frame into the main display frame, and overlay it onto the target scene image. Surface layer: ,in For the current combined layer, This indicates an overlay operation. Users can adjust the position and proportion of the character using gestures. The terminal system updates the character's state based on the gesture trajectory parameters (Δx, Δy, λ), using the following formula: ,in For gesture movement offset, For scaling ratio, For translation and scaling transformation functions;

[0179] S505: When the user selects an accessory item in the second sub-display box ∈ At that time, the terminal system applies it to the target character image in real time. The above steps involve style blending and color adjustments during rendering, using the following formula: ,in For color adjustment functions, For texture feature extraction, and For fusion weights;

[0180] S506: After the user completes the character accessory combination, the terminal system stores the current interaction state in a structured format. ,in This is the current combination state. This is the current interaction mode. For timestamps;

[0181] S507: The terminal system supports multiple characters being activated simultaneously and bound to their respective accessory sets, forming a multi-character, multi-channel mapping: Users can adjust the appearance and position of multiple characters in parallel, enabling dynamic generation of derivative images where multiple IP characters coexist in the same scene.

[0182] In this implementation scheme, the specific logical steps of step S7 are as follows:

[0183] S701: The local server receives the resource package of character images, accessory images and scene background images sent by the cloud server, and performs preprocessing operations on each image, including size standardization, format unification, background transparency and feature point extraction.

[0184] S702: The key point mapping relationship between the character and accessory images is calculated using SIFT or ORB feature matching algorithms, and the position alignment and scale correction are completed by using the affine transformation matrix T(x,y). The formula used is:

[0185] ;

[0186] in This represents the affine transformation matrix, which includes translation, rotation, and scaling parameters;

[0187] S703: Multi-resolution image decomposition is performed using the Laplacian pyramid and Gaussian pyramid, and weighted fusion is applied to each layer of the image. The formula used is as follows:

[0188] ;

[0189] in This is the image after fusion at layer l. and Laplace layers for the characters and background, respectively. and For each layer, a fusion weight mask is used to effectively smooth the stitching edges through layered fusion, maintaining consistency in details.

[0190] S704: Poisson blending is applied to the blended areas. A continuous transition between color and brightness is achieved by solving the Poisson equation. The formula used is:

[0191] ;

[0192] Where g is the source image Ω represents the target region image, and Ω represents the fusion region.

[0193] S705: Calculate the luminance and chromaticity statistics of the fused image and adjust the gain for luminance differences. ;

[0194] in This is the gain coefficient. As offset compensation parameters, histogram equalization and gamma correction are used to achieve overall style consistency and brightness balance;

[0195] S706: After fusion, an IP culture derivative image with a unified style and consistent lighting is generated. The local server compresses and encodes the resulting image and sends it to the terminal device through the local transmission protocol.

[0196] This embodiment combines speech recognition, natural language processing, semantic retrieval, and multimodal image fusion technology to achieve end-to-end automated generation from voice commands to images of cultural IP derivatives. It overcomes the problems of existing technologies, such as reliance on manual design, long generation cycles, and weak interactivity. It can simultaneously process multiple IP characters and target scenes, achieving semantic coordination and style consistency. Furthermore, through a multi-window interactive interface, it supports users to make real-time personalized adjustments to characters, accessories, and colors. Finally, the fusion algorithm generates high-quality IP cultural derivative images with consistent style and coordinated lighting, thereby significantly improving the generation efficiency, intelligence level, and interactive experience of cultural derivatives.

[0197] Case 1

[0198] (1) Voice command input and recognition stage: The user issues a voice command on the terminal device: "Generate a cherry blossom school scene with cartoon character 1 and cartoon character 2 dancing together on the playground"; The terminal device calls Tencent Cloud's voice recognition engine to recognize the voice and uses the natural language processing (NLP) module to extract the intent and generate the corresponding structured target information G={R,S,A}, where R={cartoon character 1, cartoon character 2} represents the set of IP role identifiers, S=cherry blossom school playground represents the target scene keywords, and A=dancing represents the type of command to generate. This target information is sent to the local server;

[0199] (2) Semantic parsing and retrieval stage: The local server uses the BERT model to semantically encode the target information G to obtain a semantic vector. Then, using the Faiss approximate nearest neighbor retrieval algorithm, the standard image vector sets of "cartoon character 1" and "cartoon character 2" are retrieved from the IP character database, and the scene image with the highest semantic similarity to "cherry blossom school playground" is retrieved from the scene database. The local server finally obtains a cherry blossom school scene image and two character images, and sends them to the terminal device.

[0200] (3) Related component retrieval stage: The local server retrieves the components based on the IP role identifier. A component retrieval request is sent to the cloud server, which then retrieves the accessory sets corresponding to "Cartoon Character 1" and "Cartoon Character 2" via API calls. The resource package includes: a first color change set (clothing color scheme), an accessory change set (hair accessories, backpacks, musical instruments, etc.), and a second color change set (accessory color scheme). The cloud server sends the above accessory sets back to the terminal device.

[0201] (4) Multi-window interaction stage: The terminal device displays the scene image of "Cherry Blossom School Playground" in the main display box, and displays the character images of "Cartoon Character 1" and "Cartoon Character 2" in the first sub-display box; the corresponding accessory set of the two is displayed in the second sub-display box. The user activates the image of "Cartoon Character 1" by touching it. The terminal system establishes the association channel between the character and the accessory. The user can choose to change the color of its clothing, add hair accessories or adjust the color of the decoration in the second sub-display box, and adjust the position of the character to the center of the scene by dragging it with gestures. Similarly, the user can activate the character of "Cartoon Character 2" and add accessories such as backpacks to achieve multi-character linkage editing.

[0202] (5) Image fusion and generation stage: After the user clicks the “Synthesize” button, the terminal device sends a multi-image fusion request to the local server. The local server uses multi-resolution fusion technology, Poisson fusion algorithm and gain compensation algorithm to process the image, eliminate the edge seams and color difference between the character and the scene, and output a composite image with consistent style and coordinated lighting.

[0203] (6) Output and printing stage: The terminal device displays the generated cultural IP derivative image of "Cartoon character 1 and cartoon character 2 dancing together in the cherry blossom school scene" in the main display box, and controls the printer to output physical cultural derivatives through wireless communication.

[0204] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for customizable cultural IP derivative product rapid generation, characterized in that, Comprise the following steps: S1: voice instruction recognition and target information generation: the terminal device uses Tencent cloud recognition engine and natural voice processing technology to recognize and extract the user voice instruction, generates the corresponding target information and sends it to the local server; S2: semantic analysis and multi-role scene retrieval: the local server uses the BERT model to analyze the target information in S1, and obtains the analysis result, which includes at least one IP role identifier, target scene keywords and generated instruction type. According to the analysis result, the local server performs semantic vector retrieval in the local storage IP character role library and scene library through the Faiss approximate nearest neighbor retrieval algorithm, and takes the cosine similarity and weighted fusion similarity as the matching measurement standard, and combines the image style feature consistency index for comprehensive evaluation, obtains a target role image set matched with the target information and a target scene image consistent with the target information, and sends the target scene image to the terminal device; S3: send associated accessory resource call request: the local server sends an associated component call request to the cloud server according to the IP role identifier analyzed in S2; S4: response request: the cloud server responds to the request, calls the accessory set resource package corresponding to each IP role identifier through API calling mode, and sends it to the terminal device; S5: multi-window display and interaction activation: the terminal device displays the target scene image in the main display frame, displays several target role images in the first auxiliary display frame, and displays the corresponding accessory set in the second auxiliary display frame. The user can activate any role or accessory in the auxiliary display frame through touch operation. The terminal system embedded in the terminal device automatically establishes an association channel between the role and the accessory in the main display frame after activation. After establishment, the selected role image is placed in the target scene image layer, and the role position and proportion are adjusted by gesture in the target scene image. The corresponding clothes, decorations or color replacement items in the accessory set are selected by touch, and are applied to the corresponding role image in the main display frame in real time; S6: send synthesis request: when the user completes all role and scene configuration, trigger the "synthesis" button of the terminal device, and the terminal device sends a multi-image synthesis request to the local server; S7: image synthesis processing: the local server uses multi-resolution fusion technology, Poisson fusion algorithm and gain compensation algorithm to fuse the role, accessory and scene images, eliminate the seam and color difference, generate IP cultural derivative product images with unified style and light, and send them to the terminal device; S8: output and printing: the terminal device displays the IP cultural derivative product image generated in S7 in the main display frame, and controls the printing device to perform entity printing output through wireless communication.

2. The method of claim 1, wherein, The specific logic steps of S1 are as follows: S101: the terminal device collects the user voice signal s(t) in real time through the built-in microphone array, and performs noise suppression processing on the collected original voice signal to remove the silent section and environmental noise; The formula used when using spectral subtraction for noise suppression processing is: ； wherein is the spectrum of the noise-reduced speech, is the spectrum of the original speech signal, which is a representation of the speech signal in the frequency domain obtained by performing a fast Fourier transform on s(t), reflecting the amplitude and phase of each frequency component, is the spectrum of the input speech, is the noise estimate, is the noise compensation coefficient; S102: the spectrum of the noise-reduced speech frame, window and fast Fourier transform are performed, and an acoustic feature vector is extracted wherein are all vector elements, representing the acoustic feature vector of the tthframe of speech S103: input the acoustic feature vector X into the Tencent cloud speech recognition model, calculate the acoustic probability and the language probability , and get the optimal word sequence through Viterbi decoding , and the output result is the transcription text T; The formula used above is as follows: ； where X = { }, represents an acoustic feature vector, is an acoustic feature vector of the t-th frame, is a hidden state, representing a phoneme or sub-phoneme, is an emission probability, is a state transition probability, Ta is the number of speech frames, i.e., the total number of frames after the speech signal is framed. ； where W = { represents a word sequence corresponding to the speech recognition result, is the order of the language model, is the i-th word; where P(W|X) = P(X, W) / P(X) , denotes finding the word sequence W that maximizes the joint probability, is the optimal word sequence, i.e. the sequence of characters that the terminal system finally outputs. T = ASR(X) ) = { }, where ASR(X) ) denotes an automatic speech recognition function, X denotes a sequence of acoustic features of a speech signal, T denotes a target text of a speech recognition output, i.e., a transcription text, { } denotes a sequence of words corresponding to the speech recognition result; S104: input the transcribed text T in S103 into the Tencent cloud NLP intent recognition model, perform word segmentation, part-of-speech tagging and semantic vectorization, the formula used for semantic vectorization is: ； wherein is a word vector for the ith word, is a semantic vector; The Tencent Cloud NLP intent recognition model outputs an intent category based on a classification probability. The formula used is: ; where I denotes a set of candidate intents, denotes an intent category; S105: When the intent category I * After determination, the terminal system extracts key slots from the semantic dependency tree, which includes objects, operations and attributes, and generates structured target information: ; wherein is an IP role identifier, is a target scene keyword, is a generation instruction type, is a semantic vector, is a context association parameter; S106: the terminal encapsulates the generated target information G into a JSON format message and sends it to the local server through a secure channel.

3. The method of claim 1, wherein the method further comprises: The specific logic steps of S2 are as follows: S201: The local server receives target information G from the terminal device, and inputs the text or key field in G into the BERT model for deep semantic analysis, and obtains the analysis result: PaseResult = { , , } ; wherein ; wherein represents a semantic encoding function based on the BERT model, used to map the input text into a high-dimensional semantic vector, the input text being a description of a role or a scene, represents the text description information corresponding to the i-th IP role identifier, represents the semantic feature vector of the i-th IP role text after BERT encoding, represents a target scene keyword set, represents a semantic feature vector obtained by BERT model encoding of the scene keyword.​ S202: Store the semantic vector of each image in the local IP role library and scene library respectively: RoleDB= ( ), SceneDB= ( ), wherein RoleDB represents a local IP role library, storing all role images and their corresponding semantic vectors, SceneDB represents a local scene image library, storing all scene images and their semantic vectors, represents a jth role image file, represents a semantic vector corresponding to the jth role image , represents a kth scene image file, represents a semantic vector corresponding to the kth scene image . and find the most similar vector using the Faiss Approximate Nearest Neighbor search algorithm: , ; wherein represents the i-th IP role semantic vector the role image index set with the highest similarity, represents the scene semantic vector the scene image index with the highest similarity, represents a similarity calculation function, which can be a cosine similarity, an inverse of Euclidean distance, or a weighted fusion similarity function, represents the role sample index with the highest similarity selected from the database, represents the scene sample index with the highest similarity selected from the scene library;​ S203: cosine similarity and weighted fusion similarity are used for similarity calculation, and the formula used is as follows: cosine similarity: cos_sim(A, B) = A B , wherein A represents a query target semantic vector, B represents a semantic vector of a candidate image in a database, represents a vector product point, represents the Euclidean norm of a vector, and cos_sim(A, B) represents the cosine similarity of two vectors in a semantic space, with a value range of [-1, 1]; Weighted fusion similarity: sim_total = sim_color + sim_light + sim_texture wherein and are weight coefficients, is a calculated image style consistency index, the image style consistency index of which includes color, light and texture, and sim_total represents a comprehensive similarity score for sorting and screening of candidate images. S204: sort the candidate image set, the formula used is: wherein is a sorted set of character images; Keep the first H role images consistent with the target information and high style matching degree, the formula used is: SelectedRoles = top_H(Ranked_Roles), where H is an upper limit on the number of candidates to be retained, top_H() ) denotes a function that selects the top H elements, and SelectedRoles denotes a set of H role images with the highest similarity. Perform similar screening on the scene image set, the formula used is: where SelectedScene denotes the target scene image with the highest similarity, denotes the parameter K that maximizes the objective function, denotes the comprehensive similarity between the target semantics and the scene image; S205: send the finally selected target scene image to the main display frame of the terminal device, and send the selected target role image set to the first auxiliary display frame of the terminal device.

4. The method of claim 1, wherein the method further comprises: In the S3, when sending the association component, it is necessary to send the IP role identification construction component for each IP role identification construction component call request packet: ； wherein is a current task scenario identifier, is a model feature fingerprint, used to identify the correspondence between the local matching result and the cloud resource, is a request security token, used to verify identity and prevent unauthorized calls; Before sending, the local server needs to encrypt the retrieval request packet using the HTTPS encrypted messaging protocol. The encryption method uses the following formula: Send( )= ; wherein indicates that the national secret SM4 symmetric encryption algorithm is used, is a session key, and Send() is a request transmission operation.

5. The method of claim 1, wherein, The specific logic steps of S4 are as follows: S401: The cloud server receives the association component invocation request packet from the local server After that, decryption and verification operations are performed: and legality detection is performed, Verify( )= ; wherein is a SM4 symmetric decryption function, Verify() is a verification module, used for checking the request signature and the authority, is a request security token, is a session key; S402: When the verification result is true, the cloud server determines the API interface corresponding to the resource path according to the IP role identification Determine the corresponding API interface and resource path in the resource mapping table: Wherein is a mapping function of role identification to resource path, is the interface endpoint called S403: The cloud server retrieves the IP role identifier and the corresponding accessory set resource package through API calling the corresponding database or object storage. corresponding accessory set resource package: ； wherein A is the original accessory set matched by the cloud, and B is the accessory resource library, respectively are the semantic feature vectors of the IP role and the accessory set, and sim() is a similarity calculation function: sim( , )= E( ) is a vector encoding function, is the semantic feature vector of the IP role, is the semantic feature vector of the candidate accessory set; S404: The cloud server tags and packages the retrieved accessory set constructs a resource package in a unified format: , wherein is the packaged structured accessory set, is a first color replacement set for clothes color replacement, is a decoration replacement set, is a decoration color replacement set, is model metadata containing accessory size, hierarchical relationship, and compatibility information; S405: The cloud server encrypts the structured resource package and sends it back to the terminal device through an API response, Resp( )= ( , ). After receiving it, the terminal device displays the accessory set in the second sub-display frame for user interaction.

6. The method of claim 1, wherein, The specific logic steps of S5 are as follows: S501: The terminal device receives a target scene image from a local server , a target character image set R, and a set of accessories , the following operations are performed: loading the target scene image in a main display frame , loading the target character image set R in a first auxiliary display frame , wherein are set elements, representing the nth character image in the target character image set, and loading the corresponding set of accessories in a second auxiliary display frame ={ }, wherein are set elements, representing the nth accessory material in the set of accessories; S502: the user operates the target character or accessory in the sub-display frame by touch or gesture, and the terminal system captures the touch event and determines the activation state: Active(x)= ; wherein is the activation determination threshold, Active(x)=1 indicates the activated object, and after activation, the terminal system automatically establishes a character- accessory association channel in the main display frame; S503: When the target role image is detected The terminal system establishes an association mapping when the target role image is activated: , wherein is a role-attachment mapping channel, is a set of attachments corresponding to the target role image The mapping channel is used to synchronize the replacement logic of the role in the main display frame and the attachment. S504: The terminal system will activate the target character image Drag the character image from the first sub-display frame to the main display frame and superimpose it on the target scene image The surface layer: Wherein is the current combined layer, represents the superposition operation, the user can adjust the position and scale of the character through gestures, and the terminal system updates the character state according to the gesture trajectory parameters (Δx, Δy, λ), and the formula used is: Wherein is the gesture movement offset, is the scaling ratio, is the movement and scaling transformation function; S505: When the user selects the accessory item in the second sub-display frame ∈ , the terminal system applies it to the target character image in real time , and performs style fusion and color adjustment at rendering time, using the formula: , where is the color adjustment function, is the texture feature extraction, and is the fusion weight; S506: When the user finishes the role accessory combination, the terminal system stores the current interaction state in a structured form wherein is the current combination state, is the current interaction mode, is the timestamp; S507: The terminal system supports multiple roles being activated simultaneously and being respectively bound to respective accessory sets, forming a multi-role multi-channel mapping: The user can adjust the appearance and position of multiple roles in parallel, realizing dynamic derivative image generation of multiple IP characters coexisting in the same scene.

7. The method of claim 1, wherein, The specific logic steps of S7 are as follows: S701: the local server receives the role image, accessory image and scene background image resource package sent by the cloud server, and performs preprocessing operation on each image, including size standardization, format unification, background transparency and feature point extraction; S702: adopt SIFT or ORB feature matching algorithm to calculate the key point mapping relationship between the role and accessory images, and complete position alignment and scale correction through affine transformation matrix T(x, y), the formula used is: ； wherein denotes an affine transformation matrix, containing translation, rotation and scaling parameters; S703: adopt Laplacian pyramid and Gaussian pyramid for multi-resolution image decomposition, and perform weighted fusion on each layer image, the formula used is: ； wherein is the image after fusion of the first layer, and are the Laplacian layers of the character and the background, respectively, and are the fusion weight masks corresponding to each layer, which effectively smooth the splicing edges by hierarchical fusion and maintain the consistency of details. S704: Poisson fusion processing is performed on the fusion area, and the continuous transition of color and brightness is realized by solving Poisson equation, the formula used is: ； where g is the source image, is the target region image, and Ω is the fusion region. S705: Calculate the luminance and chrominance statistics of the fused image, and perform gain adjustment on the luminance difference: ; wherein is a gain coefficient, is an offset compensation parameter, the overall style uniformity and brightness balance are realized by histogram equalization and gamma correction; S706: after fusion, generate IP cultural derivative product images with unified style and consistent lighting, the local server compresses and encodes the result image, and sends it to the terminal device through the local transmission protocol.

Images