A packaging image intelligent classification system and method based on multi-feature fusion

Abstract

The application provides a packaging image intelligent classification system and method based on multi-feature fusion, and relates to the technical fields of intelligent image classification and packaging design digitization. The application constructs a preprocessing-layered extraction-fusion coding-intelligent classification-generation linked full-link architecture; the core uses an RGBA four-channel variational autoencoder RGBAVAE to decouple packaging images with a transparent channel in layers, extracts three groups of independent latent space features of a text layer, a graphic object layer and a background layer, then uses a diffusion transformer DiT to perform deep semantic coding on the features of each layer, completes multi-feature fusion through a cross-attention mechanism, and finally realizes multi-dimensional accurate classification of industry categories, logo styles, font types and pattern elements. The classification accuracy of the application is greater than or equal to 95%, and editable layered files can be directly output, which is widely applicable to packaging material library management, design retrieval, batch generation and other scenes, and greatly improves the packaging design efficiency and the material management refinement level.

Description

Technical Field

[0001] This invention relates to the fields of intelligent image classification and digital packaging design technology, specifically to an intelligent packaging image classification system and method based on multi-feature fusion. Background Technology

[0002] With the digital transformation of the packaging design industry, enterprises and design platforms have accumulated massive amounts of packaging design materials. Traditional material management relies on manual tagging and classification, which suffers from low efficiency, inconsistent standards, and poor retrieval accuracy. Meanwhile, existing AI image classification and generation technologies have numerous technical limitations when applied to the packaging field. Classification relies on global features, resulting in coarse granularity and low accuracy: Existing mainstream image classification algorithms (such as CNN and ViT) are all based on RGB global visual feature extraction, which cannot identify the hierarchical semantics of logos, fonts, patterns, and backgrounds in packaging images; classification can only output coarse-grained categories such as "food packaging" and "daily chemical packaging", and cannot be subdivided according to logo style, font type, and pattern elements, thus failing to meet the needs of accurate retrieval of design materials.

[0003] Unable to handle transparency channels, layered feature extraction fails: Packaging design materials are mostly PNG / PSD formats with alpha transparency channels. Traditional VAEs only support RGB three-channel input, which cannot effectively encode transparency layer information; moreover, they cannot achieve layer-level decoupling of text, graphics, and background, making it difficult to extract layered semantic features for refined classification.

[0004] The classification and generation are disconnected, failing to form a closed loop: existing classification systems only output labels and cannot generate corresponding design materials; while existing generative AI-generated packaging images are mostly flat bitmaps, unable to output layered and editable PSD files, requiring designers to split them a second time, which is extremely inefficient.

[0005] Hierarchical generation technology has not been applied to classification scenarios: Existing DiT-based diffusion generation technology is mostly used for image generation, and hierarchical latent space features have not been applied to image classification tasks; hierarchical generation and intelligent classification are independent of each other and have not formed a technological synergy, so they cannot achieve a 1+1>2 effect.

[0006] Currently, there is no publicly available intelligent classification solution for packaging images that integrates RGBAVAE layer decoupling, DiT deep encoding, multi-feature fusion classification, and layered generation linkage, making it difficult to meet the digital and intelligent development needs of the packaging design industry.

[0007] Therefore, those skilled in the art provide a packaging image intelligent classification system and method based on multi-feature fusion to solve the problems mentioned in the background art. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides an intelligent classification system and method for packaging images based on multi-feature fusion, solving the following problems: 1. Solves the problems of traditional classification relying on global RGB features, inability to decouple hierarchical elements, coarse classification granularity, and low accuracy; 2. Solves the problem that traditional VAEs cannot handle alpha transparency channels and have poor accuracy in extracting layered features; 3. Resolved the issue of disconnect between classification and generation, and the inability to output editable layered files; 4. Construct a closed-loop system driven by classification to improve the efficiency of packaging material management and design; 5. Enables multi-dimensional and refined classification, supporting multi-dimensional combined searches by industry, logo, font, and pattern.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A packaging image intelligent classification system based on multi-feature fusion includes a packaging image preprocessing module, a hierarchical feature extraction module, a multi-feature fusion encoding module, an intelligent classification output module, and a hierarchical generation linkage module; The packaging image preprocessing module is used to perform format normalization, noise reduction, and channel alignment on the input packaging design image with transparency channel, and output a standardized RGBA format image. The hierarchical feature extraction module is built on the RGBA four-channel variational autoencoder RGBAVAE, which is used to decouple the standardized packaging image into three sets of independent hierarchical latent space features: text layer latent features, graphic object layer latent features, and background layer latent features. The multi-feature fusion coding module is built based on the Diffusion Transformer (DiT) and is used to perform deep semantic encoding on the three-part hierarchical latent features respectively. Then, the multi-feature fusion is completed through a cross-attention mechanism to output a global fusion feature vector. The intelligent classification output module is based on a multi-level classifier network and outputs multi-dimensional classification results according to the global fusion feature vector, which includes at least the major categories of the packaging industry, the category of logo style, the category of font type, and the category of pattern element. The layered generation linkage module is used to reverse-call DiT and RGBAVAE based on the classification results to generate editable layered packaging design images of the corresponding categories, thereby realizing a closed-loop linkage between classification and generation.

[0010] Furthermore, the RGBAVAE network includes a symmetrical encoder and decoder; The encoder maps the RGBA four-channel image to a high-dimensional latent space, and the decoder reconstructs the original RGBA image from the latent space. The training phase employs a triple constraint loss of text erasure, object erasure, and hierarchical reconstruction to force the latent space to correspond to three independent feature spaces: text semantics, graphic objects, and background texture, thereby achieving layer-level feature decoupling.

[0011] Furthermore, the DiT Diffusion Transformer employs a hierarchical conditional injection mechanism, receiving three types of hierarchical cue words: text marker CT, noise latent token, and conditional token. Text semantic conditions, graphic structure conditions, and background texture conditions are injected into different Transformer blocks, and channel-wise self-attention encoding is performed on the latent features of each layer to capture fine-grained semantic features of different layers.

[0012] Furthermore, the multi-feature fusion encoding module adopts a cross-layer attention fusion strategy: first, it performs pairwise cross-attention calculation on the encoded features of the text layer, the image layer, and the background layer to mine the semantic association between the layers; then, it uses a fully connected layer to concatenate and map the three features into a globally fused feature vector with unified dimensions, preserving the hierarchical semantics while strengthening the global association.

[0013] Furthermore, the intelligent classification output module includes a three-level classifier: the first-level classifier outputs major categories such as food, cultural and creative products, industry, and daily chemical packaging; the second-level classifier outputs sub-category tags for logo style, font type, and pattern element; and the third-level classifier outputs overall design style tags. The classification results support retrieval by single or multi-dimensional combination and automatically generate structured tags.

[0014] Furthermore, the layered generation linkage module supports two inference modes: the image-to-PSD inference mode automatically splits the input bit image into editable PSD files with text, graphics, and background layers; the text-to-PSD inference mode generates layered packaging design drawings corresponding to the categories based on natural language prompts, and the output file retains the transparency channel and layered structure.

[0015] Furthermore, a method for intelligent classification of packaging images based on multi-feature fusion includes the following steps: Step S1. Image preprocessing The input packaging image undergoes format conversion, alpha channel alignment, size normalization, and noise reduction to generate a standardized RGBA four-channel image. Step S2. Hierarchical Feature Extraction The standardized image is input into the pre-trained RGBAVAE network, and mapped to the latent space through the encoder. This decouples the text layer, the graphic object layer, and the background layer into three independent latent space features. Step S3. Hierarchical deep encoding: Input the three groups of hierarchical latent features into the DiT network, and perform deep semantic encoding respectively guided by hierarchical prompt words to output high-dimensional semantic features of each layer; Step S4. Cross-layer feature fusion A cross-attention mechanism is used to perform pairwise cross-calculation and global concatenation of the three-layer encoded features to generate a global fusion feature vector with a unified dimension. Step S5. Multi-level intelligent classification The global fusion feature vector is input into a three-level classifier network, which outputs multi-dimensional classification results of industry category, element subclass, and design style in sequence. Step S6. Generate linkage by category Based on the classification results, DiT diffusion generation and RGBAVAE decoding are reverse-driven to generate layered, editable packaging design images corresponding to the categories.

[0016] Furthermore, in step S2, the training loss function of RGBAVAE consists of three parts: image reconstruction loss, hierarchical decoupling constraint loss, and latent space regularization loss. During the training process, the hierarchical constraint loss is optimized through contrastive learning of text erasure and object erasure to ensure that the latent space features can accurately distinguish between the three types of elements: text, graphics, and background.

[0017] Furthermore, in step S3, the DiT network employs progressive hierarchical condition injection: shallow Transformer blocks inject global structural conditions, mid-layers inject graphical object conditions, and deep layers inject fine-grained text conditions; hierarchical injection enhances the semantic discriminability of features at each layer and avoids interference between features from different layers.

[0018] Furthermore, the method is applied to scenarios such as intelligent classification of material libraries, accurate retrieval of design materials, and batch style generation for food packaging, cultural and creative gift packaging, and industrial product packaging; the classification accuracy rate is ≥95%, the material retrieval efficiency is improved by more than 80%, and the generated layered files can be directly used for design editing.

[0019] This invention provides an intelligent classification system and method for packaging images based on multi-feature fusion. It has the following beneficial effects: 1. This invention provides a packaging image intelligent classification system and method based on multi-feature fusion. It applies RGBAVAE layer decoupling technology to packaging image classification, breaks through the granularity bottleneck of traditional global classification, realizes layer-level fine-grained feature extraction, and natively supports RGBA four-channel input, perfectly adapting to packaging design materials with transparent layers, filling the gap that traditional VAE cannot handle alpha channels.

[0020] 2. This invention provides a packaging image intelligent classification system and method based on multi-feature fusion. It adopts DiT hierarchical conditional injection and cross-attention multi-feature fusion, which not only preserves hierarchical semantics but also strengthens global correlation, greatly improving classification accuracy. The classification results directly drive hierarchical generation and output editable PSD files, realizing a closed-loop "classification-retrieval-generation-editing" chain.

[0021] 3. This invention provides a packaging image intelligent classification system and method based on multi-feature fusion, which can automatically label in batches, reduce the workload of manual review by 90%, improve the material retrieval efficiency by more than 80%, and the model can be deployed on cloud platforms or local workstations. It is compatible with the existing design software ecosystem and constructs a three-level classification system of industry-element-style to meet the needs of accurate retrieval and refined management of packaging design materials. Attached Figure Description

[0022] Figure 1 This is a roadmap of the RGBAVAE and DiT packaging image layering generation technology of the present invention; Figure 2 This is a schematic diagram of the overall architecture of the intelligent classification system of the present invention; Figure 3 This is an overall flowchart of the multi-feature fusion classification method of the present invention; Figure 4 This is a schematic diagram of the hierarchical feature extraction and cross-attention fusion structure of the present invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0024] In the description of this invention, unless otherwise explicitly specified and limited, the terms "connected," "linked," and "fixed" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0026] In the description of this embodiment, the terms "upper," "lower," "right," etc., refer to the orientation or positional relationship shown in the accompanying drawings. They are used only for ease of description and simplification of operation, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first" and "second" are used only for distinction in description and have no special meaning.

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0028] like Figures 1-4 As shown, this embodiment of the invention provides a packaging image intelligent classification system based on multi-feature fusion, including a packaging image preprocessing module, a hierarchical feature extraction module, a multi-feature fusion encoding module, an intelligent classification output module, and a hierarchical generation linkage module; The packaging image preprocessing module is used to perform format normalization, noise reduction, and channel alignment on the input packaging design image with transparency channel, and output a standardized RGBA format image. The hierarchical feature extraction module is built on the RGBA four-channel variational autoencoder RGBAVAE, which is used to decouple the standardized packaging image into three sets of independent hierarchical latent space features: text layer latent features, graphic object layer latent features, and background layer latent features. The multi-feature fusion coding module is built based on the Diffusion Transformer (DiT) and is used to perform deep semantic encoding on the three-part hierarchical latent features respectively. Then, the multi-feature fusion is completed through a cross-attention mechanism to output a global fusion feature vector. The intelligent classification output module is based on a multi-level classifier network and outputs multi-dimensional classification results according to the global fusion feature vector, which includes at least the major categories of the packaging industry, the category of logo style, the category of font type, and the category of pattern element. The layered generation linkage module is used to reverse-call DiT and RGBAVAE based on the classification results to generate editable layered packaging design images of the corresponding categories, thereby realizing a closed-loop linkage between classification and generation.

[0029] This invention relies on RGBAVAE to achieve layer-level feature decoupling of four-channel packaged images, uses DiT to complete deep semantic encoding of layered features, achieves multi-feature fusion through cross-attention mechanism, and finally outputs refined classification results, which in turn drive layered generation to form a full-link technology system.

[0030] The system is divided into five layers from bottom to top: Input layer: Supports multiple image formats such as bitmap, PNG, and PSD, and is compatible with design materials with alpha channels; Preprocessing layer: Performs format conversion, channel alignment, size normalization, and noise reduction, outputting a standardized RGBA four-channel image; Feature extraction coding layer: The core layer, consisting of the RGBAVAE hierarchical decoupling module and the DiT deep coding module, outputs hierarchical semantic features and global fusion features; Classification output layer: The three-level classifier outputs multi-dimensional structured labels, supporting combined retrieval; Generate a linked layer: Generate a layered and editable packaging file based on the category tags to achieve a closed loop of category-generation.

[0031] Core technology principle: RGBAVAE hierarchical feature decoupling: Traditional VAEs only process the RGB three channels. This invention expands the VAE input and output channels to four channels (R, G, B, Alpha), forming an RGBAVAE; a hierarchical constraint mechanism is introduced during the training phase: Samples with missing layers in different areas are generated by text erasure and object erasure, and paired with the original image to form a training set; The loss function incorporates a hierarchical reconstruction loss, which forces different dimensions in the latent space to correspond to text, image, and background semantics, respectively. Final implementation: Input a complete packaging image, and the encoder can output three sets of independent latent features, corresponding to the text layer, the graphic object layer, and the background layer, respectively.

[0032] DiT hierarchical deep encoding: The three-part hierarchical latent features are input into the Diffusion Transform (DiT) and a hierarchical conditional injection mechanism is adopted. Text layer features correspond to injected text marker CT conditions, enhancing the extraction of font, font size, and text semantics; The graphic layer features correspond to the structural conditions of the injected object, enhancing the shape and style features of the logo and pattern. The background layer features correspond to the injected texture conditions, enhancing the background material and color characteristics; DiT calculates high-dimensional semantic codes for each layer through self-attention and cross-attention.

[0033] Cross-attention multi-feature fusion: Employing a fusion strategy of pairwise cross-attention and global concatenation. First, calculate three sets of cross-attention features: text-image, image-background, and text-background, to uncover semantic relationships between layers; The three sets of cross features are concatenated with the original three-layer features and mapped through a fully connected layer to a global fusion feature vector of a unified dimension. The fusion feature retains the independent semantics of each layer while strengthening the correlation of the overall design style, and the classification accuracy is much higher than that of the global feature.

[0034] Three-level classification and generation linkage: The three-level classifier outputs step by step: industry category → element subcategory → style tag; The classification results can be directly used as generation conditions. Input DiT and RGBAVAEpipeline to generate layered packaging design drawings of the same style and type with one click, and output editable PSD files.

[0035] This intelligent classification method for packaged images based on multi-feature fusion includes the following steps: Step S1. Image preprocessing The input packaging image undergoes format conversion, alpha channel alignment, size normalization, and noise reduction to generate a standardized RGBA four-channel image. Step S2. Hierarchical Feature Extraction The standardized image is input into the pre-trained RGBAVAE network, and mapped to the latent space through the encoder. This decouples the text layer, the graphic object layer, and the background layer into three independent latent space features. The training loss function of RGBAVAE consists of three parts: image reconstruction loss, hierarchical decoupling constraint loss, and latent space regularization loss. During training, the hierarchical constraint loss is optimized through contrastive learning of text erasure and object erasure to ensure that the latent space features can accurately distinguish between text, graphics, and background elements. Step S3. Hierarchical deep encoding: Input the three groups of hierarchical latent features into the DiT network, and perform deep semantic encoding respectively guided by hierarchical prompt words to output high-dimensional semantic features of each layer; The DiT network employs a progressive hierarchical conditional injection: shallow Transformer blocks inject global structural conditions, mid-layers inject graphical object conditions, and deep layers inject fine-grained textual conditions; this hierarchical injection enhances the semantic discriminability of features at each layer and avoids interference between features from different layers. Step S4. Cross-layer feature fusion A cross-attention mechanism is used to perform pairwise cross-calculation and global concatenation of the three-layer encoded features to generate a global fusion feature vector with a unified dimension. Step S5. Multi-level intelligent classification The global fusion feature vector is input into a three-level classifier network, which outputs multi-dimensional classification results of industry category, element subclass, and design style in sequence. Step S6. Generate linkage by category Based on the classification results, DiT diffusion generation and RGBAVAE decoding are reverse-driven to generate layered, editable packaging design images corresponding to the categories.

[0036] This invention is implemented based on the PyTorch deep learning framework and is hardware-compatible with general-purpose GPU servers and cloud computing platforms; all model training and testing are based on real packaging design datasets.

[0037] The following sets up 3 sets of implementation examples (different application scenarios) and 1 set of comparative examples (traditional global classification method). The evaluation indicators include Top-1 classification accuracy, fine-grained element recognition accuracy, material retrieval recall rate, and generated file usability.

[0038] Example 1: Intelligent Classification System for Food Packaging Material Library ①Application Scenarios The food e-commerce design resource library contains 120,000 packaging design images in four major categories: snacks, beverages, grains and oils, and dairy products. It needs to automatically classify and label the designs and support accurate searching by logo style and font type.

[0039] ②System Configuration Preprocessing module: Supports PNG / JPG / PSD input, with a uniform size of 512×512 and four-channel normalization; RGBAVAE: 4-layer encoder and 4-layer decoder, latent space dimension 256, pre-training iterations 300,000 steps; DiT: 12-layer Transformer structure, 512 hidden dimensions, 8-head attention; Classifier: A three-level fully connected classifier that outputs 4 main categories and 32 subcategories of labels; Deployment method: The cloud-based material library has an embedded classification interface, which supports single-image and batch classification.

[0040] ③Work Process (1) Designers upload packaging materials in batches, and the preprocessing module automatically converts them into standardized RGBA images; (2) The RGBAVAE encoder decouples three sets of latent features: text, graphics, and background; (3) DiT encodes the three sets of features respectively and outputs high-dimensional semantic features; (4) Cross-attention fusion generates global features, which are input into the classifier and output as structured labels; (5) Tags are automatically written into the material library metadata, and can be searched by combination of conditions such as "national trend logo and calligraphy font".

[0041] ⑤ Test Results Top-1 accuracy in major category classification: 96.2%; Logo style recognition accuracy: 94.8%; Font type recognition accuracy: 95.5%; Material retrieval recall rate: 93.7%; Batch processing speed: 1000 images / 5 minutes.

[0042] Example 2: Layered Classification and Generation Linkage System for Cultural and Creative Packaging ①Application Scenarios A cultural and creative gift design platform needs to classify and manage existing materials, and at the same time support generating layered design drafts of the same style with one click based on the classification results, and outputting editable PSD files.

[0043] ②System Configuration Based on Example 1, a layered generation linkage module is added; Image to PSD: Input a bitmap image, automatically split it into three layers: text, graphics, and background, and output a PSD file with an alpha channel; Text to PSD: Input the text prompt "Chinese trend mooncake packaging" to generate a layered design image of the corresponding category; The generated files are compatible with mainstream design software such as Photoshop and Figma.

[0044] ③ Running effect Image to PSD layering accuracy: 92.3%; The generated design drawing matched the target classification with a accuracy of 91.5%. The average revision time for designers has been reduced from 2 hours to 40 minutes, improving design efficiency by 67%. Material retrieval efficiency improved by 85%.

[0045] Example 3: Industrial Packaging Batch Classification Management System ①Application Scenarios Large packaging companies' archives management departments need to classify and archive 500,000 historical industrial product packaging drawings in batches and establish a digital archive database.

[0046] ②System Configuration Optimize the batch processing interface to support batch task queues of tens of thousands; The three-level classification is simplified to two levels: industry category and product subcategory. It includes a tag export function that automatically generates an Excel categorized ledger.

[0047] ③ Running effect Category classification accuracy: 93.7%; Processing time for a single batch of 10,000 images: 42 minutes; Manual review workload reduced by 91%; The archiving period has been shortened from 3 months to 15 days.

[0048] Comparative Example 1: Traditional ResNet50 global image classification method We used the mainstream ResNet50 convolutional neural network, extracting only RGB global visual features, and trained and classified the same dataset of 120,000 food packaging images.

[0049] Performance testing: The data of the systems in Examples 1-3 and Comparative Example 1 were tested, and the test results are shown in the table below:

[0050] Based on the above test comparison results, it can be seen that 1. This invention adopts a classification scheme of hierarchical decoupling and multi-feature fusion, which improves the accuracy of major categories by 15 percentage points and the accuracy of fine-grained element recognition by more than 30 percentage points, with extremely obvious advantages; 2. Traditional methods can only achieve coarse classification, while this invention supports fine-grained classification, hierarchical file splitting, and targeted generation, forming a complete closed loop; 3. Customized RGBA four-channel processing and layered decoupling for packaging design scenarios perfectly adapt to the characteristics of industry materials, and its practical value far exceeds that of general classification algorithms.

[0051] This invention is developed based on a general deep learning framework and can be deployed in various forms such as cloud servers, local workstations, and design platform plugins, adapting to the needs of enterprises of different sizes. This invention supports incremental training on user-owned private datasets and can be quickly adapted to different sub-sectors such as food, daily chemicals, cultural and creative products, and industry. This invention requires no dedicated hardware and can run using existing GPU computing power. It can be directly embedded into existing design material libraries, e-commerce platforms, and design software ecosystems. This invention can be widely applied to packaging design companies, brand owners, e-commerce platforms, and material trading platforms, covering a wide range of industries and possessing high commercial value.

[0052] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A packaging image intelligent classification system based on multi-feature fusion, characterized in that, It includes a packaging image preprocessing module, a hierarchical feature extraction module, a multi-feature fusion encoding module, an intelligent classification output module, and a hierarchical generation linkage module; The packaging image preprocessing module is used to perform format normalization, noise reduction, and channel alignment on the input packaging design image with transparency channel, and output a standardized RGBA format image. The hierarchical feature extraction module is built on the RGBA four-channel variational autoencoder RGBAVAE, which is used to decouple the standardized packaging image into three sets of independent hierarchical latent space features: text layer latent features, graphic object layer latent features, and background layer latent features. The multi-feature fusion coding module is built based on the Diffusion Transformer (DiT) and is used to perform deep semantic encoding on the three-part hierarchical latent features respectively. Then, the multi-feature fusion is completed through a cross-attention mechanism to output a global fusion feature vector. The intelligent classification output module is based on a multi-level classifier network and outputs multi-dimensional classification results according to the global fusion feature vector, which includes at least the major categories of the packaging industry, the category of logo style, the category of font type, and the category of pattern element. The layered generation linkage module is used to reverse-call DiT and RGBAVAE based on the classification results to generate editable layered packaging design images of the corresponding categories, thereby realizing a closed-loop linkage between classification and generation.

2. The intelligent classification system for packaging images based on multi-feature fusion according to claim 1, characterized in that, The RGBAVAE network includes a symmetrical encoder and decoder; The encoder maps the RGBA four-channel image to a high-dimensional latent space, and the decoder reconstructs the original RGBA image from the latent space. The training phase employs a triple constraint loss of text erasure, object erasure, and hierarchical reconstruction to force the latent space to correspond to three independent feature spaces: text semantics, graphic objects, and background texture, thereby achieving layer-level feature decoupling.

3. The intelligent classification system and method for packaging images based on multi-feature fusion according to claim 1, characterized in that, The DiT Diffusion Transformer employs a hierarchical conditional injection mechanism, receiving three types of hierarchical cue words: text marker CT, noise latent token, and conditional token. Text semantic conditions, graphic structure conditions, and background texture conditions are injected into different Transformer blocks, and channel-wise self-attention encoding is performed on the latent features of each layer to capture fine-grained semantic features of different layers.

4. The intelligent classification system and method for packaging images based on multi-feature fusion according to claim 1, characterized in that, The multi-feature fusion encoding module adopts a cross-layer attention fusion strategy: first, it performs pairwise cross-attention calculation on the encoded features of the text layer, the image layer, and the background layer to mine the semantic association between the layers; then, it uses a fully connected layer to concatenate and map the three features into a global fusion feature vector with unified dimensions, which preserves the hierarchical semantics while strengthening the global association.

5. The intelligent classification system for packaging images based on multi-feature fusion according to claim 1, characterized in that, The intelligent classification output module includes a three-level classifier: the first-level classifier outputs major categories such as food, cultural and creative products, industry, and daily chemical packaging; the second-level classifier outputs sub-category tags for logo style, font type, and pattern element; and the third-level classifier outputs overall design style tags. The classification results support retrieval by single or multi-dimensional combination and automatically generate structured tags.

6. The intelligent classification system for packaging images based on multi-feature fusion according to claim 1, characterized in that, The layered generation linkage module supports two inference modes: the image-to-PSD inference mode automatically splits the input image into editable PSD files with text, graphics, and background layers; the text-to-PSD inference mode generates layered packaging design drawings corresponding to the categories based on natural language prompts, and the output file retains the transparency channel and layered structure.

7. A method for intelligent classification of packaging images based on multi-feature fusion, characterized in that, Includes the following steps: Step S1. Image preprocessing The input packaging image undergoes format conversion, alpha channel alignment, size normalization, and noise reduction to generate a standardized RGBA four-channel image. Step S2. Hierarchical Feature Extraction The standardized image is input into the pre-trained RGBAVAE network, and mapped to the latent space through the encoder. This decouples the text layer, the graphic object layer, and the background layer into three independent latent space features. Step S3. Hierarchical deep encoding: Input the three groups of hierarchical latent features into the DiT network, and perform deep semantic encoding respectively guided by hierarchical prompt words to output high-dimensional semantic features of each layer; Step S4. Cross-layer feature fusion A cross-attention mechanism is used to perform pairwise cross-calculation and global concatenation of the three-layer encoded features to generate a global fusion feature vector with a unified dimension. Step S5. Multi-level intelligent classification The global fusion feature vector is input into a three-level classifier network, which outputs multi-dimensional classification results of industry category, element subclass, and design style in sequence. Step S6. Generate linkage by category Based on the classification results, DiT diffusion generation and RGBAVAE decoding are reverse-driven to generate layered, editable packaging design images corresponding to the categories.

8. The intelligent classification method for packaging images based on multi-feature fusion according to claim 7, characterized in that, In step S2, the training loss function of RGBAVAE consists of three parts: image reconstruction loss, hierarchical decoupling constraint loss, and latent space regularization loss. During training, the hierarchical constraint loss is optimized through comparative learning of text erasure and object erasure to ensure that the latent space features can accurately distinguish between text, graphics, and background elements.

9. The intelligent classification method for packaging images based on multi-feature fusion according to claim 7, characterized in that, In step S3, the DiT network employs progressive hierarchical condition injection: shallow Transformer blocks inject global structural conditions, mid-layers inject graphical object conditions, and deep layers inject fine-grained text conditions; hierarchical injection enhances the semantic discriminability of features at each layer and avoids interference between features from different layers.

Images