Packaging design intelligent generation and optimization integrated system and method

Abstract

This invention discloses an integrated system and method for intelligent generation and optimization of packaging design, comprising: a multi-stage generation module for sequentially executing layout generation, element generation, text rendering, and style optimization to generate an initial packaging design, wherein the text rendering adopts a text generation model based on the Transformer architecture; an intelligent evaluation module for calculating a weighted sum as a comprehensive evaluation score based on five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability; and a reinforcement learning optimization module for constructing a reinforcement learning model based on a deep deterministic policy gradient algorithm. This invention achieves fully automated generation and multi-dimensional closed-loop optimization of packaging design, significantly shortening the design cycle, improving design quality, text rendering accuracy, compliance, and manufacturability, and possessing adaptability to different packaging categories.

Description

Technical Field

[0001] This invention belongs to the field of packaging design technology, and in particular relates to an integrated system and method for intelligent generation and optimization of packaging design. Background Technology

[0002] Packaging design is a crucial part of product development, involving collaborative work across multiple dimensions such as layout, graphic elements, text typography, and color scheme. It must also meet brand visual identity standards, industry regulations, and printing production process constraints. Currently, packaging design primarily relies on manual drawing and adjustments using professional design software (such as Adobe Illustrator and CorelDRAW). Some design platforms offer template libraries and semi-automated layout assistance functions, such as rule-based or case-based layout recommendations and preset color schemes. In recent years, with the development of deep learning technology, methods for generating graphic elements based on generative adversarial networks or diffusion models have emerged, capable of generating illustrations or textures from text descriptions. Furthermore, some research attempts to combine optical character recognition with rule engines for automatic text layout on packaging, and to use pre-trained neural network models for aesthetic scoring or brand consistency detection of design drafts. These technologies alleviate repetitive work for designers to some extent, but they remain in the stage of scattered, modular auxiliary tools, and have not yet formed an integrated solution covering the entire process.

[0003] However, existing packaging design technologies still have the following shortcomings in practical applications: First, the design process is fragmented, with layout, elements, text, and style each independent of each other, lacking a closed-loop linkage mechanism from generation to optimization, resulting in long design cycles that often require weeks of repeated revisions. Second, design quality is highly dependent on human experience; the works produced by different designers vary significantly in terms of aesthetics, brand consistency, and readability, and are difficult to quantify and consistently guarantee. Third, text rendering accuracy is insufficient; traditional methods, such as OCR post-processing or font library matching, cannot achieve industrial-grade precision in character spacing, glyphs, and alignment that perfectly integrates with the layout. Fourth, compliance checks rely on manual verification of mandatory labeling information, font copyrights, and industry warnings, which is prone to omissions, especially in pharmaceutical and hazardous materials packaging with high regulatory requirements. Fifth, manufacturability verification is delayed; problems such as insufficient bleed, incorrect color modes, and excessively thin lines often only become apparent after the design drafts are delivered to the printing plant, leading to frequent rework. Sixth, there is a lack of adaptive optimization capabilities for different packaging categories; fixed-weight evaluation models struggle to accommodate the differentiated needs of different scenarios such as food, cosmetics, and electronic products. Seventh, existing methods are mostly sequential and independent optimizations, unable to make cross-stage global collaborative adjustments to layout, elements, text, and style, and are prone to getting trapped in local optima. Therefore, there is an urgent need for a system and method that can achieve integrated intelligent generation and optimization of packaging design, and has multi-dimensional comprehensive evaluation and reinforcement learning closed-loop iteration capabilities. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of existing technologies, this invention provides an integrated system and method for intelligent generation and optimization of packaging design, which solves the problems of long design cycles, reliance on experience for quality, low text rendering accuracy, easy omission of compliance, lagging manufacturability, and lack of adaptive and global collaborative optimization in existing technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution: An integrated system and method for intelligent generation and optimization of packaging design, including: The multi-stage generation module is used to sequentially execute four stages: layout generation, element generation, text rendering, and style optimization to generate the initial packaging design. Among them, the text rendering stage uses a text generation model based on the Transformer architecture to convert text strings into text images in the packaging layout. The intelligent evaluation module is used to obtain the packaging design generated by the multi-stage generation module or reinforcement learning optimization module, and calculate scores based on five dimensions: aesthetics, brand consistency, readability, compliance and manufacturability. Then, the weighted sum of the scores of each dimension is calculated as the comprehensive evaluation score. The reinforcement learning optimization module is used to construct a reinforcement learning model based on the deep deterministic policy gradient algorithm. It uses the comprehensive evaluation score output by the intelligent evaluation module as the reward signal, the feature vector of the packaging design as the state space, and the adjustment amount of the generation parameters of each stage in the multi-stage generation module as the action space. The action is output through the policy network and applied to the multi-stage generation module to iteratively generate the optimized packaging design until the comprehensive evaluation score is greater than or equal to a preset threshold or the number of iterations reaches a preset maximum value.

[0006] Preferably, the state space includes: layout feature vector, element feature vector, text feature vector, style encoding vector, and a multi-dimensional vector composed of the scores of the five dimensions; the action space includes the adjustment amount of at least one of the following generation parameters: layout grid density, element position offset, element scaling ratio, text font size, text line spacing, global hue value, and global contrast value; the weight coefficients of each dimension in the weighted sum are determined according to the packaging design type.

[0007] Preferably, the text generation model based on the Transformer architecture includes: An encoder is used to receive text strings and typographic constraints, and output text semantic features; A decoder is used to generate a text image based on the semantic features of the text through a self-attention mechanism, wherein the glyphs, character spacing, and alignment of the text image conform to the typographic constraints.

[0008] Preferably, in the intelligent evaluation module: The aesthetic score is output by a machine learning model; Brand consistency score is obtained by calculating the similarity between the primary color scheme and auxiliary graphics in the packaging design and the pre-set brand visual identity specifications. The readability score includes evaluations of text contrast, minimum font size, and background clutter. The compliance score includes the results of checks on the completeness of mandatory labeling information on packaging, font copyright compliance, and the correctness of industry warnings; The manufacturability score includes the results of testing whether the packaging design meets printing process constraints, which include at least bleed size, cut marks, color mode, and minimum line width.

[0009] The preferred method for intelligent generation and optimization of packaging design includes the following steps: S1: Multi-stage generation steps: The four sub-steps of layout generation, element generation, text rendering, and style optimization are executed sequentially to generate the initial packaging design; among them, the text rendering sub-step uses a text generation model based on the Transformer architecture to convert text strings into text images; S2: Intelligent evaluation steps: Obtain the current packaging design, calculate the scores for the five dimensions of aesthetics, brand consistency, readability, compliance and manufacturability, and then calculate the weighted sum of the scores for each dimension as the comprehensive evaluation score; S3: Reinforcement learning optimization steps: Construct a reinforcement learning model based on a deep deterministic policy gradient algorithm, using the feature vector of the current packaging design as the state space, the adjustment amount of the generation parameters of each sub-step as the action space, and the comprehensive evaluation score as the reward signal. Output the action through the policy network, and update the generation parameters of at least one sub-step according to the action to generate the optimized packaging design. S4: Closed-loop iteration step: Repeat S2 and S3 until the comprehensive evaluation score is greater than or equal to the preset threshold or the number of iterations reaches the preset maximum value, and output the final packaging design.

[0010] Preferably, in step S3, the Actor network and the Critic network are trained based on a deep deterministic policy gradient algorithm; the Actor network outputs a deterministic action based on the current state, which corresponds to an adjustment of a set of generation parameters; the Critic network evaluates the cumulative reward that can be obtained by performing the action in the current state; during training, a quadruple of historical state, action, reward and next state is randomly sampled from the experience replay storage area to update the parameters of the Critic network and the Actor network respectively.

[0011] Preferably, in step S1, the layout generation sub-step outputs the layout grid structure and information partitions of the packaging; the element generation sub-step generates patterns, illustrations, or background textures within the layout grid structure; and the style optimization sub-step adjusts the overall tone, contrast, sharpness, and texture of the packaging design.

[0012] Preferably, in step S1, during the text rendering sub-step, the Transformer model also receives compliance check rules as constraints, adjusts the forced annotation text to meet the minimum font size and contrast requirements, and replaces it with a compliant similar font when non-compliant font copyright is detected.

[0013] Preferably, in step S4, the iteration is terminated when the difference between the comprehensive evaluation scores of two consecutive iterations is less than a preset difference threshold.

[0014] Preferably, in S1, the layout generation sub-step, element generation sub-step, text rendering sub-step, and style optimization sub-step are sequentially connected, and the input of the subsequent sub-step is the output of the previous sub-step; in S3, the action output by the reinforcement learning optimization step includes parallel adjustments to multiple sub-steps, simultaneously adjusting at least two parameters among layout, element, text, and style.

[0015] The technical effects and advantages of the integrated system and method for intelligent generation and optimization of packaging design of this invention are as follows: 1. This invention achieves fully automated packaging design generation by sequentially executing four stages—layout generation, element generation, text rendering, and style optimization—through a multi-stage generation module, eliminating the need for manual step-by-step operations. Simultaneously, through a reinforcement learning optimization module based on a closed-loop iterative framework constructed using a deep deterministic policy gradient algorithm, the system can autonomously optimize design parameters and quickly reach preset thresholds. Compared to traditional manual design processes, this invention compresses the packaging design cycle from weeks to hours, significantly improving design efficiency.

[0016] 2. This invention uses an intelligent evaluation module to comprehensively evaluate packaging design from five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability, and calculates a weighted sum as the overall evaluation score. This five-dimensional evaluation system covers all key quality indicators of packaging design, avoiding the limitations of optimizing a single indicator. The reinforcement learning optimization module uses the overall evaluation score as a reward signal, outputs actions through a policy network, and iteratively optimizes, ensuring that the final design meets preset standards in multiple dimensions, resulting in a significantly higher overall quality than traditional manual design.

[0017] 3. This invention employs an end-to-end text generation model based on the Transformer architecture in the text rendering stage. This model receives text strings and typographic constraints through an encoder, and the decoder generates text images that meet the requirements for glyphs, character spacing, and alignment through a self-attention mechanism. This architecture can directly map text semantic features into high-quality text images in the packaging layout, ensuring industrial-grade accuracy and solving the problems of poor text rendering effects and inaccurate typography in traditional methods.

[0018] 4. The compliance score of this invention includes the results of checks on the completeness of mandatory labeling information on packaging, font copyright compliance, and the correctness of industry warnings. The reinforcement learning optimization module can automatically adjust the generated parameters based on the compliance score, supplement missing mandatory labeling information, adjust the text to meet minimum font size and contrast requirements, and replace non-compliant fonts with compliant fonts when non-compliant font copyrights are detected. This mechanism effectively avoids legal risks caused by regulatory omissions and is particularly suitable for packaging design of pharmaceuticals, hazardous materials, and other products with high compliance requirements.

[0019] 5. The manufacturability score of this invention includes the detection results of bleed size, trim marks, color mode, and minimum line width. The system can automatically identify and correct design parameters that do not meet printing process constraints, and output production files that can be directly delivered to printing, reducing repeated communication and modifications between the design draft and the printing plant, and lowering the production error rate.

[0020] 6. In this invention, the weighting coefficients of the weighted sum are determined according to the packaging design type. The state space contains the feature vectors of the packaging design, and the action space contains the adjustment amounts of the generated parameters at each stage. The system can automatically adjust the weights and optimization strategies of the evaluation dimensions for different product categories (such as food, cosmetics, pharmaceuticals, and electronic products), and has wide applicability.

[0021] 7. In this invention, the four sub-steps of the multi-stage generation module are sequentially connected, with the input of the subsequent sub-step being the output of the previous sub-step, ensuring the continuity of the design process. The actions output by the reinforcement learning optimization module include parallel adjustments for multiple sub-steps, capable of simultaneously adjusting at least two parameters among layout, elements, text, and style, achieving cross-stage global collaborative optimization and avoiding the local optima problem of independent step-by-step optimization. Attached Figure Description

[0022] Figure 1 This is a system framework diagram of the integrated system and method for intelligent generation and optimization of packaging design proposed in this invention; Figure 2 This is a flowchart of the integrated system and method for intelligent generation and optimization of packaging design proposed in this invention. Detailed Implementation

[0023] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0025] refer to Figure 1-2This invention provides an integrated system and method for intelligent generation and optimization of packaging design, comprising a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module. The multi-stage generation module sequentially executes four stages: layout generation, element generation, text rendering, and style optimization, to generate an initial packaging design. The text rendering stage uses a Transformer-based text generation model to convert text strings into text images on the packaging layout. The intelligent evaluation module calculates scores based on five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability, and calculates a weighted sum as a comprehensive evaluation score. The reinforcement learning optimization module constructs a reinforcement learning model based on a deep deterministic policy gradient algorithm, using the comprehensive evaluation score as a reward signal, the feature vector of the packaging design as the state space, and the adjustment amounts of the generation parameters in each stage of the multi-stage generation module as the action space. Actions are output through a policy network and applied to the multi-stage generation module, iteratively generating an optimized packaging design until a stopping condition is met. This invention achieves fully automated generation and multi-dimensional comprehensive optimization of packaging design from layout to style through the deep integration of multi-stage generation and reinforcement learning closed-loop optimization. It significantly shortens the design cycle, improves design quality and compliance, and has the ability to adapt to different packaging categories.

[0026] System overall architecture: This invention provides an integrated intelligent generation and optimization system for packaging design, comprising: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0027] The multi-stage generation module sequentially executes four stages: layout generation, element generation, text rendering, and style optimization, to generate the initial packaging design. Specifically, the multi-stage generation module includes: a layout generation submodule, used to generate the packaging's grid structure and information partitions; an element generation submodule, used to generate patterns, illustrations, or background textures within the grid structure; a text rendering submodule, used to convert text strings into text images on the packaging layout using a Transformer-based text generation model; and a style optimization submodule, used to adjust the overall tone, contrast, sharpness, and texture of the packaging design. The text rendering submodule uses a Transformer-based text generation model consisting of an encoder and a decoder: the encoder receives the text string and typographic constraints, outputting text semantic features; the decoder generates a text image based on these semantic features using a self-attention mechanism, ensuring that the glyphs, character spacing, and alignment of this text image conform to the typographic constraints.

[0028] The intelligent evaluation module is used to acquire packaging designs generated by the multi-stage generation module or reinforcement learning optimization module, and calculates scores based on five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability. The weighted sum of these scores is then used as the comprehensive evaluation score. Specifically, the intelligent evaluation module includes: an aesthetics evaluation submodule, implemented by a machine learning model, which outputs an aesthetics score; a brand consistency evaluation submodule, which calculates the similarity between the main color scheme, auxiliary graphics, and preset brand visual identity specifications in the packaging design to obtain a brand consistency score; a readability evaluation submodule, which calculates the evaluation results of text contrast, minimum font size, and background interference as a readability score; a compliance evaluation submodule, which checks the completeness of mandatory labeling information on the packaging, font copyright compliance, and the correctness of industry warnings, and outputs a compliance score; and a manufacturability evaluation submodule, which checks whether the packaging design meets printing process constraints and outputs a manufacturability score. Printing process constraints include at least bleed size, trim marks, color mode, and minimum line width.

[0029] The reinforcement learning optimization module is used to construct a reinforcement learning model based on a deep deterministic policy gradient algorithm. It uses the comprehensive evaluation score output by the intelligent evaluation module as the reward signal, the feature vector of the packaging design as the state space, and the adjustment amounts of the generation parameters in each stage of the multi-stage generation module as the action space. Actions are output through the policy network and applied to the multi-stage generation module, iteratively generating optimized packaging designs until the comprehensive evaluation score is greater than or equal to a preset threshold or the number of iterations reaches a preset maximum value. Specifically, the reinforcement learning optimization module includes: an Actor network, used to output deterministic actions based on the current state, each action corresponding to a set of adjustment amounts for generation parameters; a Critic network, used to evaluate the cumulative reward obtained by performing the action in the current state; an experience replay storage area, used to store a quadruple of historical states, actions, rewards, and the next state; and a training unit, used to randomly sample quadruples from the experience replay storage area and update the parameters of the Critic network and the Actor network respectively. The state space includes: layout feature vectors, element feature vectors, text feature vectors, style encoding vectors, and a multi-dimensional vector composed of five dimensions of scores. The action space includes adjustments to at least one of the following generation parameters: layout grid density, element position offset, element scaling, text font size, text line spacing, global hue value, and global contrast value. The weighting coefficients for each dimension in the weighted sum are determined based on the packaging design type. Example

[0030] Purpose of implementation: This embodiment aims to verify the basic performance of the present invention in the food packaging design scenario, especially the ability of the multi-stage generation module to generate the initial packaging design, the five-dimensional comprehensive evaluation capability of the intelligent evaluation module, and the ability of the reinforcement learning optimization module to improve the comprehensive evaluation score to a preset threshold through closed-loop iteration.

[0031] Implementation System: The integrated intelligent generation and optimization system for packaging design according to the present invention includes: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0032] The multi-stage generation module includes: a layout generation submodule, used to generate the layout grid structure and information partitions of the packaging; an element generation submodule, used to generate patterns, illustrations or background textures within the layout grid structure; a text rendering submodule, used to convert text strings into text images in the packaging layout using a Transformer-based text generation model, which includes 6 Transformer encoding layers and 6 Transformer decoding layers, each containing 8 attention heads, with a feedforward network dimension of 2048; and a style optimization submodule, used to adjust the overall tone, contrast, sharpness and texture of the packaging design.

[0033] The intelligent evaluation module includes: an aesthetic evaluation submodule, which uses a ResNet-50 regression model pre-trained on the AVA dataset; a brand consistency evaluation submodule, which uses K-means clustering (k=3) to extract the main color tone, edge detection and template matching to identify auxiliary graphics, and calculates the cosine similarity with the preset brand visual recognition specifications; a readability evaluation submodule, which calculates text contrast (based on the brightness difference between the foreground and background colors), minimum font size (converted to physical font size through pixel height), and background interference (based on the texture complexity within the text area); a compliance evaluation submodule, which uses keyword matching to detect the completeness of mandatory annotation information, uses font metadata to verify font copyright compliance, and uses regular expressions to match the correctness of industry warnings; and a productivity evaluation submodule, which detects bleed size (the outward expansion of design edges), crop marks (crosshairs and registration marks), color mode (whether it is CMYK), and minimum line width (whether it is greater than the printing lower limit).

[0034] The reinforcement learning optimization module includes: an Actor network, a 3-layer fully connected neural network (input layer 256 → hidden layer 512 → output layer 10), used to output a deterministic action based on the current state, which corresponds to the adjustment amount of a set of generation parameters; a Critic network, a 3-layer fully connected neural network (input layer (state + action) 512 → hidden layer 256 → output layer 1), used to evaluate the cumulative reward that can be obtained by performing the action in the current state; an experience replay storage area with a capacity of 10,000 quadruplets; and a training unit that randomly samples 64 quadruplets each time and updates the target network parameters using a soft update method (τ=0.001).

[0035] Implementation steps: S1. Multi-stage generation steps: The multi-stage generation module sequentially executes four sub-steps: layout generation, element generation, text rendering, and style optimization, to generate the initial packaging design.

[0036] In the layout generation sub-step, the layout generation sub-module receives the size information of the food packaging (a square box of 200mm×150mm×50mm), calls the layout generation model based on conditional generative adversarial network, outputs a grid system of 4 rows and 6 columns, and divides the packaging surface into the main visual area, product name area, nutrition information area, barcode area and net content area.

[0037] In the element generation sub-step, the element generation sub-module uses a pattern generator based on a diffusion model to generate a circular main visual illustration based on the user-input keywords "orange, juice, fresh". It also selects auxiliary graphics (orange slice texture) that match the brand color scheme from the preset material library and lays them flat on the background area.

[0038] In the text rendering sub-step, the text rendering module adopts a text generation model based on the Transformer architecture. The encoder of this model receives the text string "freshly squeezed orange juice" and typographic constraints (font: Source Han Sans Bold, font size: 24pt to 36pt, line spacing: 1.2 times the font size, alignment: center alignment, maximum text area boundary: 120mm × 40mm). The decoder generates a text image through a self-attention mechanism, whose glyphs, character spacing, and alignment conform to the typographic constraints.

[0039] In the style optimization sub-step, the style optimization module adjusts the overall tone, contrast, sharpness, and texture of the packaging design. Specifically, tone adjustment is achieved through hue rotation, contrast adjustment through S-curve mapping, sharpness adjustment through unsharpened masks, and texture adjustment through overlaying matte paper textures, generating the initial packaging design.

[0040] S2. Intelligent Evaluation Steps: The intelligent evaluation module obtains the current packaging design and calculates scores in five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability.

[0041] The aesthetics assessment submodule scored 0.78. The brand consistency assessment submodule calculated a cosine similarity of 0.92. The readability assessment submodule scored 0.71. The compliance assessment submodule scored 0.6 due to the detection of a missing "net content" label. The manufacturability assessment submodule scored 0.95 because the bleed size, trim marks, color mode, and minimum line width all met the requirements.

[0042] The intelligent evaluation module calculates the weighted sum of the scores for each dimension as the comprehensive evaluation score. In this embodiment, based on the packaging design type (juice beverage packaging), the weighting coefficients are set as follows: aesthetics 0.3, brand consistency 0.25, readability 0.2, compliance 0.15, and manufacturability 0.1, resulting in a comprehensive evaluation score of 0.791.

[0043] S3. Reinforcement learning optimization steps: The reinforcement learning optimization module constructs a reinforcement learning model based on the deep deterministic policy gradient algorithm. It uses the feature vector of the current packaging design as the state space, the adjustment amount of the generation parameters of each sub-step as the action space, and the comprehensive evaluation score as the reward signal. It outputs the action through the policy network and updates the generation parameters of at least one sub-step according to the action to generate the optimized packaging design.

[0044] In this embodiment, the state space includes layout feature vectors, element feature vectors, text feature vectors, style encoding vectors, and a multi-dimensional vector composed of five-dimensional scores. The action space includes adjustment values ​​for parameters such as layout grid density, element position offset, element scaling, text font size, text line spacing, global hue value, and global contrast value. Specifically, the Actor network outputs actions such as: increasing net text content, changing white text to dark orange to improve contrast, and fine-tuning the main visual position.

[0045] S4. Closed-loop iteration steps: Repeat steps S2 and S3 until the overall evaluation score is greater than or equal to a preset threshold or the number of iterations reaches a preset maximum value. In this embodiment, the preset threshold is 0.85, and the maximum number of iterations is 30. After the first iteration, the overall evaluation score increases to 0.84, and after the second iteration, it reaches 0.87, meeting the stopping condition, and the final packaging design is output. Implementation effect

[0046] This embodiment successfully generated a compliant juice beverage packaging design. The initial comprehensive evaluation score was 0.791, which improved to 0.87 after two reinforcement learning iterations, reaching the preset threshold. The entire design cycle was approximately 2.5 hours, while traditional manual design typically takes several weeks. The intelligent evaluation module successfully identified issues in the initial design such as the lack of net content labeling and insufficient text contrast. The reinforcement learning optimization module automatically corrected these defects by adjusting the corresponding parameters in the action space. The final packaging design performed well in all five dimensions: aesthetics, brand consistency, readability, compliance, and manufacturability. Example

[0047] Purpose of implementation: This embodiment aims to verify the applicability of the present invention in the high-end cosmetic packaging scenario, especially the effect of the style optimization sub-step on improving the aesthetic score, and the impact of adjusting the weight coefficients under different packaging design types on the comprehensive evaluation results.

[0048] Implementation System: The integrated intelligent generation and optimization system for packaging design according to the present invention includes: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0049] The multi-stage generation module includes a layout generation submodule, an element generation submodule, a text rendering submodule, and a style optimization submodule, with the same structure as in Example 1. In this example, the Transformer model configuration of the text rendering submodule is consistent with that in Example 1, and the style optimization submodule is additionally configured with a metallic sheen effect generation function.

[0050] The intelligent evaluation module includes: an aesthetic evaluation submodule, a brand consistency evaluation submodule, a readability evaluation submodule, a compliance evaluation submodule, and a manufacturability evaluation submodule, with the same structure as in Example 1. In this example, based on the packaging design type (high-end skincare products), the weighting coefficients are preset as follows: aesthetics 0.4, brand consistency 0.3, readability 0.1, compliance 0.1, and manufacturability 0.1.

[0051] The reinforcement learning optimization module includes an Actor network, a Critic network, an experience replay storage area, and training units, with the same structure as in Example 1. In this example, since the initial comprehensive evaluation score has reached the threshold, the reinforcement learning optimization module does not actually participate in the iteration.

[0052] Implementation steps: S1. Multi-stage generation steps: The multi-stage generation module receives cosmetic packaging size information (80mm×80mm×50mm face cream box). The layout generation submodule outputs a 3x3 grid system, with information partitions including the main visual area (70%), brand logo area (15%), ingredient information area (10%), and barcode area (5%). The element generation submodule generates a product rendering image as the main visual. The text rendering submodule uses a Transformer-based text generation model to render the brand name "LUXESKIN," with typographic constraints including Didot font, font size range of 18pt to 22pt, and serif font. The generated text image has a metallic sheen effect. The style optimization submodule adjusts the hue, contrast, sharpness, and texture: the hue is adjusted to rose gold and matte black, the contrast is reduced to present a soft effect, the sharpness is set to medium, and a frosted glass texture is overlaid on the texture.

[0053] S2. Intelligent Evaluation Steps: The intelligent evaluation module obtains the current packaging design and calculates scores for five dimensions. The aesthetic evaluation submodule outputs 0.91, the brand consistency evaluation submodule outputs 0.95, the readability evaluation submodule outputs 0.88, the compliance evaluation submodule outputs 1.0 (ingredient list includes all INCI names, net content is fully labeled), and the manufacturability evaluation submodule outputs 0.92 (color mode CMYK, minimum line width 0.2mm, bleed 2.5mm). The weighted sum yields a comprehensive evaluation score of 0.85.

[0054] S3. Reinforcement learning optimization steps: Since the overall evaluation score has reached the preset threshold of 0.85, the action output of the Actor network in the reinforcement learning optimization module is a zero vector (no adjustment is needed).

[0055] S4. Closed-loop iteration steps: If the overall evaluation score is detected to be greater than or equal to the preset threshold, the iteration is terminated and the current packaging design is output directly. Implementation effect

[0056] In this embodiment, the initial packaging design achieved a comprehensive evaluation score of 0.85, directly meeting the stopping condition without requiring reinforcement learning iterations. The design cycle was approximately 1.8 hours. This demonstrates that the present invention can adaptively adjust the evaluation weights according to different packaging design types. For cosmetic packaging with high requirements for aesthetics and brand consistency, the initial design generated by the multi-stage generation module already possesses high quality. The style optimization submodule's fine adjustments to color tone and material texture significantly improved the aesthetic score, while the metallic text and frosted glass texture ensured that the final design aligns with the brand positioning of high-end skincare products. Example

[0057] Purpose of implementation: This embodiment aims to verify the invention's ability to prioritize compliance in the pharmaceutical packaging scenario, particularly whether the reinforcement learning optimization module can supplement the mandatory labeling information and adjust the text to meet regulatory requirements through multiple iterations when the compliance score is low.

[0058] Implementation System The integrated intelligent generation and optimization system for packaging design according to the present invention includes: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0059] The structure of the multi-stage generation module is the same as in Example 1. In this example, the text rendering submodule is additionally configured with specific layout constraints for pharmaceutical packaging, including a lower limit for the generic name number of the drug (2.5mm), and requirements for the color and position of warning messages.

[0060] The structure of the intelligent assessment module is the same as in Example 1. In this example, based on the packaging design type (over-the-counter drug), the weighting coefficients are preset as follows: aesthetics 0.15, brand consistency 0.10, readability 0.20, compliance 0.45, and manufacturability 0.10. The compliance assessment submodule is configured with 15 mandatory information check rules, including OTC identification, approval number, production date, expiration date, storage conditions, contraindications, etc.

[0061] The structure of the reinforcement learning optimization module is the same as in Example 1. In this example, the preset threshold is 0.9, and the maximum number of iterations is 50. The compliance dimension has a high weight in the reward function.

[0062] Implementation steps: S1. Multi-stage generation steps: The multi-stage generation module receives the packaging size information for over-the-counter cold medicine (a 120mm × 80mm × 20mm folded cardboard box). The layout generation submodule outputs the layout grid structure. In the text rendering submodule, the Transformer-based text generation model receives layout constraints, including that the generic drug name number must be no less than 2.5mm, and the warning "Please take as directed" must be in red and placed in a prominent position. The multi-stage generation module generates text images that conform to the above constraints. In the initial design, information such as drug name, approval number, and production date were generated, but the "contraindications" section was missing.

[0063] S2. Intelligent Evaluation Steps: The intelligent evaluation module obtains the current packaging design and calculates scores for five dimensions. The compliance evaluation submodule detected a missing "prohibited" item, resulting in a compliance score of 0.4. The scores for the other dimensions are: Aesthetics 0.65, Brand Consistency 0.70, Readability 0.75, and Manufacturability 0.85. The weighted sum yields a comprehensive evaluation score of 0.55.

[0064] S3. Reinforcement learning optimization steps: The reinforcement learning optimization module constructs a reinforcement learning model with a high weight for compliance scores in the state space. The Actor network outputs the following actions: adding a "Contraindicated" text area (located below the ingredient list), adding a "Not for use by pregnant women" warning, adjusting the warning font size to 2.8mm, and increasing the warning contrast to 7:1.

[0065] S4. Closed-loop iteration steps: Repeat steps S2 and S3. After the first iteration, the compliance score increases to 0.65; after the fifth iteration, it increases to 0.85; after the twelfth iteration, the compliance score reaches 0.98, and the overall evaluation score reaches 0.91, meeting the stopping condition and outputting the final packaging design. Implementation effect

[0066] In this embodiment, the initial packaging design lacked a "contraindication" section, resulting in a compliance score of only 0.4 and an overall evaluation score of 0.55, far below the threshold of 0.9. After 12 iterations of reinforcement learning optimization, the system automatically supplemented the missing mandatory labeling information and adjusted the font size and contrast of the warning statements, ultimately improving the compliance score to 0.98 and the overall evaluation score to 0.91. The entire design cycle was approximately 4.2 hours, while traditional pharmaceutical packaging design typically requires more than two weeks due to regulatory complexity. This embodiment verifies the significant advantages of the present invention in scenarios with high compliance requirements; the reinforcement learning optimization module can effectively drive the system to fill regulatory gaps and avoid legal risks. Example

[0067] Purpose of implementation: This embodiment aims to verify the ability of the present invention to optimize manufacturability in the context of electronic product packaging, especially when the initial design does not meet printing process constraints (such as incorrect color mode, insufficient minimum line width, lack of bleed and trim marks), and whether the reinforcement learning optimization module can automatically correct these defects.

[0068] Implementation System: The integrated intelligent generation and optimization system for packaging design according to the present invention includes: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0069] The structure of the multi-stage generation module is the same as that in Example 1.

[0070] The structure of the intelligent evaluation module is the same as in Example 1. In this example, based on the packaging design type (electronic product packaging), the weighting coefficients are preset as follows: Aesthetics 0.25, Brand Consistency 0.20, Readability 0.15, Compliance 0.10, and Manufacturability 0.30. The Manufacturability evaluation submodule is specifically configured with bleed size detection (requires at least 2mm), cut mark detection (crosshairs and registration marks), color mode detection (requires CMYK), and minimum line width detection (requires greater than 0.1mm).

[0071] The structure of the reinforcement learning optimization module is the same as in Example 1. In this example, the action space is specifically expanded to include discrete action items: converting the color mode to CMYK, adding bleed, adding clipping markers, etc.

[0072] Implementation steps: S1. Multi-stage generation steps: The multi-stage generation module receives the phone packaging box dimensions (180mm×90mm×20mm, cardboard mounting). The layout generation submodule outputs the page grid structure. The element generation submodule generates the product rendering and brand logo. The text rendering submodule generates the product name and specifications. The style optimization submodule adjusts the overall color scheme. In the initial design, the color mode was RGB (not meeting printing requirements), the minimum line width was 0.08mm (less than the minimum requirement for the printing press), and no bleed or trim marks were added.

[0073] S2. Intelligent Evaluation Steps: The intelligent evaluation module obtains the current packaging design and calculates scores across five dimensions. The manufacturability evaluation submodule detected an incorrect color mode (RGB instead of CMYK), insufficient minimum line width, and missing bleed and trim marks, resulting in a manufacturability score of 0.2. The scores for the other dimensions are: Aesthetics 0.75, Brand Consistency 0.80, Readability 0.85, and Compliance 0.90. The weighted sum yields a comprehensive evaluation score of 0.41.

[0074] S3. Reinforcement learning optimization steps: The reinforcement learning optimization module constructs a reinforcement learning model. The action space includes adjustments such as converting the color mode to CMYK, increasing bleed, adding clipping markers, and thickening thin lines. The Actor network outputs the following actions: converting the color mode from RGB to CMYK, increasing the bleed by 3mm, adding crosshairs and registration marks as clipping markers, and increasing the minimum line width to 0.12mm.

[0075] S4. Closed-loop iteration steps: Repeat S2 and S3. After the first iteration, the productivity score increases to 0.70; after the third iteration, it increases to 0.90; after the fifth iteration, the productivity score reaches 0.96, and the overall evaluation score reaches 0.89, meeting the stopping condition, and the final packaging design is output. Implementation effect

[0076] In this embodiment, the initial packaging design performed reasonably well in terms of aesthetics and brand consistency, but its overall evaluation score was only 0.41 due to manufacturability defects. After five iterations of reinforcement learning optimization, the system automatically corrected all manufacturability issues, including color mode, bleed size, trim marks, and line width, ultimately improving the overall evaluation score to 0.89. The final output design file can be directly delivered to the printing plant for production without manual modification. The design cycle is approximately 3 hours, while in traditional processes, repeated communication and modifications between designers and printing plants typically take several days. This embodiment verifies the adaptive capability of the present invention to printing process constraints. Example

[0077] Purpose of implementation: This embodiment aims to verify the adaptability of the present invention to various packaging categories, in particular whether the system can automatically adjust the evaluation weights according to different packaging design types, and sequentially process three significantly different packaging tasks—beverage bottle labels, chocolate gift boxes, and pesticide bottles—on the same system.

[0078] Implementation System: The integrated intelligent generation and optimization system for packaging design according to the present invention includes: a multi-stage generation module, an intelligent evaluation module, and a reinforcement learning optimization module.

[0079] The structure of the multi-stage generation module is the same as in Example 1. The structure of the intelligent evaluation module is the same as in Example 1, and it includes a built-in mapping table between product categories and weight coefficients, supporting automatic weight configuration for various types of packaging, such as beverage packaging, food gift boxes, and hazardous material packaging. The structure of the reinforcement learning optimization module is the same as in Example 1.

[0080] In this embodiment, the system processes three tasks sequentially, and the weight coefficient of each task is automatically adjusted according to the packaging design type.

[0081] Implementation steps: Users input three packaging design tasks into the system in sequence: Task ① is for beverage bottle labels, Task ② is for chocolate gift boxes, and Task ③ is for pesticide bottles (dangerous goods).

[0082] Task 1 (Beverage Bottle Label): S1. The multi-stage generation module generates an initial design based on the beverage packaging type.

[0083] S2. The intelligent evaluation module identifies the packaging design type as beverage packaging and sets the weight coefficients as follows: aesthetics 0.30, brand consistency 0.25, readability 0.20, compliance 0.15, and manufacturability 0.10. The overall evaluation score is calculated to be 0.72.

[0084] S3. The reinforcement learning optimization module outputs optimized actions.

[0085] S4. After three closed-loop iterations, the comprehensive evaluation score is improved to 0.86, reaching the preset threshold, and the final design is output.

[0086] Task 2 (Chocolate Gift Box): S1. The multi-stage generation module generates an initial design based on the type of food gift box packaging.

[0087] S2. The intelligent evaluation module identifies the packaging design type as food gift box packaging, sets the weight coefficients as aesthetics 0.35, brand consistency 0.30, readability 0.15, compliance 0.10, and manufacturability 0.10, and calculates the comprehensive evaluation score as 0.78.

[0088] S3. The reinforcement learning optimization module outputs optimized actions.

[0089] S4. After two closed-loop iterations, the comprehensive evaluation score is improved to 0.88, reaching the preset threshold, and the final design is output.

[0090] Task 3 (Pesticide Bottle): S1. The multi-stage generation module generates an initial design based on the type of hazardous materials packaging, while the text rendering submodule places particular emphasis on generating hazardous materials warning statements.

[0091] S2: The intelligent assessment module identifies the packaging design type as hazardous material packaging and sets weighting coefficients as follows: Aesthetics 0.10, Brand Consistency 0.10, Readability 0.20, Compliance 0.50, and Manufacturability 0.10. It focuses on checking mandatory information such as hazardous material warnings ("Highly Toxic," "Keep Out of Children"), pesticide registration number, and production license number. The compliance assessment submodule initially outputs 0.3, with a comprehensive assessment score of 0.35.

[0092] S3. In the reward function of the reinforcement learning optimization module, the compliance dimension has a high weight. The Actor network output actions include supplementing missing warning messages, adjusting the position of warning messages to a prominent area, and increasing the font size of warning messages.

[0093] S4. After 8 closed-loop iterations, the compliance score improved to 0.95, and the comprehensive evaluation score reached 0.88, meeting the stopping conditions, and the final packaging design was output. Implementation effect

[0094] In this embodiment, the same system successfully handled three significantly different packaging design tasks. For beverage bottle labels and chocolate gift boxes, aesthetics and brand consistency had high weights; for pesticide bottles, compliance had a weight as high as 0.5, with the system prioritizing the completeness and prominence of hazard warnings. The system can automatically adjust the weight coefficients according to the packaging design type without manual intervention. In task ③, the initial compliance score was only 0.3, which improved to 0.95 after 8 iterations of reinforcement learning optimization, fully verifying the reliability of the invention in high-risk product categories. The entire process is fully automated, with design cycles of 2.2 hours, 1.9 hours, and 5.1 hours respectively, significantly better than traditional manual design.

[0095] Comparative Example 1 Purpose of implementation: This comparative example aims to verify the technical advantages of the present invention in terms of packaging design cycle, design quality, and automation by comparing it with the system of the present invention. The comparative example adopts a traditional manual design process and does not include the multi-stage generation module, intelligent evaluation module, and reinforcement learning optimization module of the present invention.

[0096] Implementation System: The comparison model does not use the system of this invention, but instead employs traditional manual design tools (Adobe Illustrator) and a manual review process. Specifically, this includes: designers using Adobe Illustrator for layout design, illustration, text typesetting, and style adjustments; manual compliance checks against a regulatory checklist; and manual communication with the printing plant to confirm manufacturability requirements.

[0097] Implementation steps: The design task is the same as in Example 1 (juice beverage packaging, 200mm×150mm×50mm square box).

[0098] Step 1 (analogous to S1): A designer with 3 years of experience manually creates an artboard based on the design brief provided by the brand, sets bleed and crop marks, and divides the layout. This takes approximately 2 hours. Next, the designer uses the pen tool to draw an orange illustration and background texture, revising multiple times before finalizing the design. This takes approximately 6 hours. Then, the designer types in text such as "freshly squeezed orange juice," selects the font "Source Han Sans Bold," adjusts the font size and line spacing, and manually aligns the text. This takes approximately 1.5 hours. Finally, the designer adjusts parameters such as hue and contrast, tries various color schemes, and determines the final style. This takes approximately 2 hours.

[0099] Step Two (analogous to S2): The designer checks the mandatory labeling information against food labeling regulations, and adds any missing net content labeling. This takes approximately 0.5 hours.

[0100] Step 3 (analogous to S3): The designer modifies the design file according to the printing plant's requirements, confirming bleed, trim marks, color mode, etc., but the trim marks were missed and will be added after feedback from the printing plant. This takes about 1 hour.

[0101] Step Four (analogous to S4): The designer delivers the final design documents to the brand for confirmation. If there are any revisions, the above steps are repeated. The actual project cycle is usually 3 to 5 business days.

[0102] The entire design process took approximately 13 hours (excluding communication and waiting time). Implementation effect

[0103] In the comparative study, the final packaging design documents produced by manual design were evaluated by the intelligent evaluation module of this invention (used only for post-design evaluation and not involved in the design process), and the overall evaluation score was 0.74. Specifically, the aesthetic score was 0.76, the brand consistency score was 0.80, the readability score was 0.70 (slightly lower contrast), the compliance score was 0.85 (initially missing net content but subsequently added), and the manufacturability score was 0.65 (deduction due to missing cutting marks). The design cycle was approximately 13 hours (excluding communication waiting time) and required designers to possess significant professional experience and regulatory knowledge.

[0104] Compared with Example 1: The design cycle of Example 1 of this invention is 2.5 hours, which is about one-fifth of the comparative example; the comprehensive evaluation score is 0.87, which is about 17.6% higher than the comparative example. Furthermore, this invention is fully automated, requiring no human intervention, while manual design relies on personal experience and is prone to overlooking compliance requirements and manufacturability details. This comparative example verifies the significant technical advantages of this invention in terms of efficiency, quality, and automation.

[0105] Compared to Examples 1-5 and Comparative Example 1, Comparative Example 1 of this invention employs a traditional manual design process. Designers manually complete layout, element drawing, text typesetting, style adjustment, and compliance checks using tools such as Adobe Illustrator. The entire design process takes approximately 13 hours (excluding communication and waiting time with brand owners and printing plants), and the actual project cycle is typically 3 to 5 working days. In contrast, the system in each embodiment of this invention operates fully automatically: Example 1 (food packaging) undergoes two reinforcement learning iterations, with a design cycle of approximately 2.5 hours; Example 2 (cosmetic packaging) meets the initial design requirements, with a cycle of approximately 1.8 hours; Example 3 (pharmaceutical packaging) has high compliance requirements, undergoing 12 iterations, with a cycle of approximately 4.2 hours; Example 4 (electronic product packaging) undergoes 5 iterations to correct manufacturability defects, with a cycle of approximately 3 hours; Example 5 (multi-category adaptive design) handles beverages, food gift boxes, and pesticide bottles in only 2.2 hours, 1.9 hours, and 5.1 hours, respectively. In contrast, the design cycle of each embodiment of the present invention is only about one-fifth to one-tenth of that of the comparative example, and no human intervention is required throughout the process, with a degree of automation far exceeding that of traditional manual design.

[0106] In Comparative Example 1, the final packaging designed manually was evaluated by the intelligent evaluation module of this invention, and the overall evaluation score was 0.74. Specifically, the readability score was 0.70 (low text contrast), the compliance score was 0.85 (initially missing net content labeling, still incomplete after manual supplementation), and the manufacturability score was 0.65 (missing cutting marks, leading to rework from the printing plant). In contrast, the final outputs of each embodiment of this invention reached the preset thresholds (0.85-0.9): Example 1's overall evaluation score increased from the initial 0.791 to 0.87; Example 2 directly reached 0.85; Example 3 increased from 0.55 to 0.91; Example 4 increased from 0.41 to 0.89; and Example 5 achieved scores of 0.86, 0.88, and 0.88 for each task, respectively. Particularly in the dimensions of compliance (pharmaceutical packaging increased from 0.4 to 0.98) and manufacturability (electronic product packaging increased from 0.2 to 0.96), this invention automatically corrects defects through reinforcement learning closed-loop iteration, avoiding omissions and rework that are prone to occur in manual design. In summary, this invention not only reduces the design cycle from several weeks to hours, but also improves the overall design quality by about 15%-20%, with particularly outstanding advantages in scenarios with high compliance and high manufacturability requirements.

[0107] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0108] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for intelligent generation and optimization of packaging design, characterized in that, include: The multi-stage generation module is used to sequentially execute four stages: layout generation, element generation, text rendering, and style optimization to generate the initial packaging design. Among them, the text rendering stage uses a text generation model based on the Transformer architecture to convert text strings into text images in the packaging layout. The intelligent evaluation module is used to obtain the packaging design generated by the multi-stage generation module or reinforcement learning optimization module, and calculate scores based on five dimensions: aesthetics, brand consistency, readability, compliance and manufacturability. Then, the weighted sum of the scores of each dimension is calculated as the comprehensive evaluation score. The reinforcement learning optimization module is used to construct a reinforcement learning model based on the deep deterministic policy gradient algorithm. It uses the comprehensive evaluation score output by the intelligent evaluation module as the reward signal, the feature vector of the packaging design as the state space, and the adjustment amount of the generation parameters of each stage in the multi-stage generation module as the action space. The action is output through the policy network and applied to the multi-stage generation module to iteratively generate the optimized packaging design until the comprehensive evaluation score is greater than or equal to a preset threshold or the number of iterations reaches a preset maximum value.

2. The integrated system of intelligent generation and optimization of packaging design according to claim 1, wherein, The state space includes: layout feature vector, element feature vector, text feature vector, style encoding vector, and a multi-dimensional vector composed of the scores of the five dimensions; the action space includes the adjustment amount of at least one of the following generation parameters: layout grid density, element position offset, element scaling ratio, text font size, text line spacing, global hue value, and global contrast value; the weight coefficients of each dimension in the weighted sum are determined according to the packaging design type.

3. The integrated system of claim 1, wherein, The text generation model based on the Transformer architecture includes: An encoder is used to receive text strings and typographic constraints, and output text semantic features; A decoder is used to generate a text image based on the semantic features of the text through a self-attention mechanism, wherein the glyphs, character spacing, and alignment of the text image conform to the typographic constraints.

4. The integrated system of claim 1, wherein, In the intelligent evaluation module: The aesthetic score is output by a machine learning model; Brand consistency score is obtained by calculating the similarity between the primary color scheme and auxiliary graphics in the packaging design and the pre-set brand visual identity specifications. The readability score includes evaluations of text contrast, minimum font size, and background clutter. The compliance score includes the results of checks on the completeness of mandatory labeling information on packaging, font copyright compliance, and the correctness of industry warnings; The manufacturability score includes the results of testing whether the packaging design meets printing process constraints, which include at least bleed size, cut marks, color mode, and minimum line width.

5. The packaging design intelligent generation and optimization integrated system of packaging design intelligent generation and optimization integrated method according to any one of claims 1-4, characterized in that, Includes the following steps: S1: Multi-stage generation steps: The four sub-steps of layout generation, element generation, text rendering, and style optimization are executed sequentially to generate the initial packaging design; among them, the text rendering sub-step uses a text generation model based on the Transformer architecture to convert text strings into text images; S2: Intelligent evaluation steps: Obtain the current packaging design, calculate the scores for the five dimensions of aesthetics, brand consistency, readability, compliance and manufacturability, and then calculate the weighted sum of the scores for each dimension as the comprehensive evaluation score; S3: Reinforcement learning optimization steps: Construct a reinforcement learning model based on a deep deterministic policy gradient algorithm, using the feature vector of the current packaging design as the state space, the adjustment amount of the generation parameters of each sub-step as the action space, and the comprehensive evaluation score as the reward signal. Output the action through the policy network, and update the generation parameters of at least one sub-step according to the action to generate the optimized packaging design. S4: Closed-loop iteration step: Repeat S2 and S3 until the comprehensive evaluation score is greater than or equal to the preset threshold or the number of iterations reaches the preset maximum value, and output the final packaging design.

6. The method of claim 5, wherein the method is integrated with the optimization of the packaging design. In S3, the Actor network and the Critic network are trained based on the deep deterministic policy gradient algorithm; the Actor network outputs a deterministic action based on the current state, which corresponds to the adjustment amount of a set of generation parameters; The Critic network evaluates the cumulative reward that can be obtained by performing the action in the current state; during training, a quadruple of historical state, action, reward and next state is randomly sampled from the experience replay storage area to update the parameters of the Critic network and the Actor network respectively.

7. The method of claim 5, wherein the method is integrated with the optimization of the packaging design. In step S1, the layout generation sub-step outputs the layout grid structure and information partitions of the packaging; the element generation sub-step generates patterns, illustrations or background textures within the layout grid structure; and the style optimization sub-step adjusts the overall tone, contrast, sharpness and texture of the packaging design.

8. The method of claim 5, wherein the method is integrated with the optimization of the packaging design. 8 In S1, during the text rendering sub-step, the Transformer model also receives compliance check rules as constraints, forces the labeled text to be adjusted to meet the minimum font size and contrast requirements, and replaces it with a compliant similar font when it detects that the font copyright is not compliant.

9. The method of claim 5, wherein the method is integrated with the optimization of the packaging design.

9. The method of claim 5, wherein the method is integrated with the optimization of the packaging design. In step S4, the iteration is terminated when the difference between the comprehensive evaluation scores of two consecutive iterations is less than a preset difference threshold.

10. The method of claim 5, wherein the method is integrated with the optimization of the packaging design. In S1, the layout generation sub-step, element generation sub-step, text rendering sub-step, and style optimization sub-step are sequentially connected, with the input of the subsequent sub-step being the output of the previous sub-step; in S3, the action output by the reinforcement learning optimization step includes parallel adjustments to multiple sub-steps, simultaneously adjusting at least two parameters among layout, element, text, and style. ​

Images