A garbage classification method and system based on intelligent analysis

Abstract

The application belongs to the field of intelligent classification, and specifically relates to a garbage classification method and system based on intelligent analysis, which comprises a visual feature extraction module, a structured semantic knowledge base construction module, a semantic prior generation module, a classifier training module and a garbage classification module; the application introduces double-flow coding of local texture and global shape into a semantic enhancement framework, and realizes adaptive fusion through a learnable gate to dynamically adjust the degree of dependence on the two kinds of features; the application proposes a four-dimensional semantic ontology as a semantic framework for garbage classification, and designs a two-level routing network to realize dimension-aware expert decomposition, so that hierarchical design makes semantic modeling of garbage classification more fine; through retrieval and optimization of the dimension gate, the application ensures that the retrieved garbage classification text and the sample are highly matched in important dimensions, and further enhances the consistency and class discriminability of cross-view representation through comparison of semantic consistency loss, so as to finally improve the garbage classification efficiency.

Description

Technical Field

[0001] This invention belongs to the field of intelligent classification, specifically a waste classification method and system based on intelligent analysis. Background Technology

[0002] Waste sorting is a major environmental governance issue facing modern society. With the development of deep learning technology, computer vision-based automatic waste sorting systems have been applied in some scenarios. However, existing automatic systems suffer from low sorting efficiency due to issues such as differences in the scale of visual features, differences in the structure of semantic knowledge, and differences in the definition of retrieval relevance. Summary of the Invention

[0003] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a waste classification method and system based on intelligent analysis. Addressing the problem of low classification efficiency caused by differences in visual feature scale, semantic knowledge structure, and retrieval relevance definitions, this invention introduces dual-stream encoding of local texture and global shape into a semantic enhancement framework. Through learnable gating, adaptive fusion is achieved, enabling the model to dynamically adjust its dependence on these two features based on the visual state of the waste (intact / damaged, clear / occluded), thereby improving classification accuracy. This invention proposes a four-dimensional semantic ontology as the semantic framework for waste classification and designs a two-level routing network to achieve dimension-aware expert decomposition. The first-level routing determines dimensional importance, and the second-level routing assigns experts within each dimension. This hierarchical design makes the semantic modeling of waste classification more refined. Through dimensional gating retrieval and optimization, this invention introduces dimension-aware similarity measurement, ensuring a high degree of matching between retrieved waste classification text and samples in important dimensions. Furthermore, by comparing semantic consistency loss, it further enhances the consistency and class discriminative power of cross-perspective representations, improving waste classification efficiency.

[0004] This invention provides a waste sorting method based on intelligent analysis, which specifically includes the following steps:

[0005] Step S1: Obtain images of the waste items to be classified and extract their visual features;

[0006] The step of extracting its visual features further includes:

[0007] Step S11: Use parallel local texture encoders and global shape encoders to extract the local texture feature matrix and global shape feature vector of the garbage image, respectively;

[0008] Step S12: Calculate the fusion weights through an adaptive gating network, and perform weighted fusion of the pooled representation of local texture features and the global shape feature vector according to the fusion weights to generate a multi-scale visual summary vector, i.e., visual features.

[0009] The adaptive gating network calculates the fusion weights. The formula is: ;

[0010] In the formula, For local texture feature matrix, This is an image-level local representation obtained by average pooling along the local region dimension; This is the global shape feature vector after linear projection alignment; This represents a vector concatenation operation; It is a learnable gating weight matrix; This is a gated bias term; Use the Sigmoid activation function; For fusion weights;

[0011] The multi-scale visual summarization vector The calculation formula is: ;

[0012] Step S2: Construct a structured semantic knowledge base for the field of waste classification, wherein the semantic knowledge base contains textual description features with multiple preset semantic dimensions;

[0013] The steps for constructing a structured semantic knowledge base for the field of waste sorting include:

[0014] Step S21: Define a semantic ontology for waste classification that includes the dimensions of material properties, shape and structure, usage category, and disposal method of waste;

[0015] Step S22: For each waste category, use a large language model to generate specific descriptive text with multiple semantic dimensions according to each semantic ontology;

[0016] Step S23: Encode the specific descriptive text of each semantic dimension into text description features through a text encoder to form a structured semantic knowledge base corresponding to the category, and calculate the dimension prototype vector for each semantic dimension. The dimension prototype vector is used to represent the semantic centroid of the semantic dimension in the feature space.

[0017] Step S3: Decompose the text description features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism, and generate sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues.

[0018] The step of decomposing textual descriptive features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism includes:

[0019] Step S31: Construct a dimensional decomposition semantic refinement module, which includes multi-dimensional expert groups corresponding to each semantic dimension, and each dimension expert group contains several expert query matrices;

[0020] Step S32: Set up the first-level routing network to calculate the dimensional importance weight vector of each semantic dimension based on the multi-scale visual summarization vector: ;

[0021] In the formula, This refers to the multi-scale visual summarization vector. This is the first-level routing weight matrix, used to map visual features to a dimensionality importance space. This is the first-level routing bias vector. Represents the activation function. Represents the dimension importance weight vector;

[0022] Step S33: Calculate the routing score for each expert based on textual description features within each semantic dimension using the second-level routing network: ;

[0023] In the formula, Represents textual description features. This represents a linear projection function specific to dimension a. The Gaussian noise term representing input dependency is used to prevent routing network collapse and promote diversity and specialization in expert learning. The expert routing score matrix is ​​represented by dimension a;

[0024] Step S34: Based on the routing score, select some experts for each semantic dimension, and use the expert query matrix corresponding to the selected experts to decompose the text description features into fine-grained semantic cues corresponding to the dimensions and experts.

[0025] The steps for generating sample-level semantic priors include:

[0026] Step S35: Calculate multi-head cross attention using the fine-grained semantic clues as the query and the local texture feature matrix as the key and value;

[0027] Step S36: Incorporate the dimensional importance weights into the cross-attention calculation to adaptively enhance the attention response of specific semantic dimensions based on the visual characteristics of the current sample;

[0028] Step S37: Weight and fuse the expert outputs in each semantic dimension according to the expert routing score to obtain the fused representation of each dimension;

[0029] Step S38: The fusion representations of each dimension are weighted and fused according to the importance weight of the dimensions to obtain a category-level semantic representation;

[0030] Step S39: Aggregate the category-level semantic representations of all categories to obtain the sample-level semantic prior;

[0031] Step S4: Based on the sample-level semantic prior, retrieve matching multi-perspective text features from the structured semantic knowledge base, generate auxiliary semantic prior using the retrieved multi-perspective text features, and train the classifier model by combining the main prior and auxiliary prior through a random optimization strategy.

[0032] The steps for retrieving matching multi-perspective text features from a structured semantic knowledge base include:

[0033] Step S41: Align the sample-level semantic priors in the feature space to obtain adapted prior features;

[0034] Step S42: Calculate the dimension-aware comprehensive similarity measure based on the dimension importance weights, the similarity between the adapted prior features and the prototype vectors of each dimension, and the similarity between the text description features and the prototype vectors of each dimension.

[0035] Step S43: Based on the comprehensive similarity metric, select several text description features with the highest similarity from the structured semantic knowledge base of the corresponding category to form a retrieval set;

[0036] Step S44: Randomly shuffle the retrieval set and store it in a queue so that the random optimization strategy can retrieve it in a random order;

[0037] The steps of generating auxiliary semantic priors using retrieved multi-view text features and training a classifier by jointly using the main prior and auxiliary priors through a random optimization strategy include:

[0038] Step S45: In each training iteration, pop a text description feature from the queue;

[0039] The pop-up text description features are decomposed into auxiliary fine-grained semantic cues through the decompositional semantic refinement mechanism, and auxiliary semantic priors are generated through visual-text cross-attention interaction.

[0040] Step S46: Input the sample-level semantic prior and the auxiliary semantic prior into the shared classifier to obtain the main classification result and the auxiliary classification result, and then fuse the two to obtain the final classification result;

[0041] Step S47: Calculate the classification loss based on the final classification result and the true label, and update the model parameters with the classification loss as the optimization objective;

[0042] Step S5: Input the garbage image to be classified into the trained model to obtain the garbage classification result.

[0043] The present invention provides a waste sorting system based on intelligent analysis, including a visual feature extraction module, a structured semantic knowledge base construction module, a semantic prior generation module, a classifier training module, and a waste sorting module;

[0044] The visual feature extraction module acquires images of the waste items to be classified and extracts their visual features.

[0045] The structured semantic knowledge base construction module constructs a structured semantic knowledge base for the field of waste classification. The semantic knowledge base contains text description features with multiple preset semantic dimensions.

[0046] The semantic prior generation module uses a decompositional semantic refinement mechanism to decompose textual descriptive features into multiple fine-grained semantic cues, and generates sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues.

[0047] The classifier training module, based on the sample-level semantic prior, retrieves matching multi-perspective text features from the structured semantic knowledge base, generates auxiliary semantic priors using the retrieved multi-perspective text features, and trains the classifier model by combining the main prior and auxiliary priors through a random optimization strategy.

[0048] The waste sorting module inputs images of waste to be sorted into the trained model to obtain waste sorting results.

[0049] The beneficial results achieved by the present invention using the above solution are as follows:

[0050] (1) This invention introduces dual-stream encoding of local texture and global shape into the semantic enhancement framework, and achieves adaptive fusion through learnable gating, so that the model can dynamically adjust the dependence on the two features according to the visual state of the garbage (intact / damaged, clear / occluded), thereby improving the classification accuracy;

[0051] (2) This invention proposes a four-dimensional semantic ontology as the semantic framework for garbage classification, and designs a two-level routing network to realize dimension-aware expert decomposition. The first-level routing determines the importance of the dimension, and the second-level routing assigns experts within the dimension. The hierarchical design makes the semantic modeling of garbage classification more refined.

[0052] (3) This invention introduces dimension-aware similarity measurement through dimension-gated retrieval and optimization, ensuring that the retrieved garbage classification text and samples are highly matched in important dimensions, and further enhances the consistency of cross-perspective representation and category discrimination by comparing semantic consistency loss, thereby improving the efficiency of garbage classification. Attached Figure Description

[0053] Figure 1 A block diagram of a waste sorting system based on intelligent analysis provided by the present invention;

[0054] Figure 2 This is a flowchart illustrating a waste sorting method based on intelligent analysis provided by the present invention.

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0056] 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.

[0057] Example 1, see Figure 2 The present invention provides a waste sorting method based on intelligent analysis, which specifically includes the following steps:

[0058] Step S1: Obtain images of the waste items to be classified and extract their visual features;

[0059] Step S2: Construct a structured semantic knowledge base for the field of waste classification, wherein the semantic knowledge base contains textual description features with multiple preset semantic dimensions;

[0060] Step S3: Decompose the text description features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism, and generate sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues.

[0061] Step S4: Based on the sample-level semantic prior, retrieve matching multi-perspective text features from the structured semantic knowledge base, generate auxiliary semantic prior using the retrieved multi-perspective text features, and train the classifier model by combining the main prior and auxiliary prior through a random optimization strategy.

[0062] Step S5: Input the garbage image to be classified into the trained model to obtain the garbage classification result.

[0063] Example 2: This example is based on the above examples, and the step of extracting its visual features includes:

[0064] Step S11: Use parallel local texture encoders and global shape encoders to extract the local texture feature matrix and global shape feature vector of the garbage image, respectively;

[0065] Step S12: Calculate the fusion weights through an adaptive gating network, and perform weighted fusion of the pooled representation of local texture features and the global shape feature vector according to the fusion weights to generate a multi-scale visual summary vector, i.e., visual features.

[0066] Example 3: This example is based on the above examples, and the adaptive gating network calculates the fusion weights. The formula is: ;

[0067] In the formula, For local texture feature matrix, This is an image-level local representation obtained by average pooling along the local region dimension; This is the global shape feature vector after linear projection alignment; This represents a vector concatenation operation; It is a learnable gating weight matrix; This is a gated bias term; Use the Sigmoid activation function; For fusion weights;

[0068] The multi-scale visual summarization vector The calculation formula is: .

[0069] Example 4: Based on the above examples, the steps for constructing a structured semantic knowledge base for the field of waste sorting include:

[0070] Step S21: Define a semantic ontology for waste classification that includes the dimensions of material properties, shape and structure, usage category, and disposal method of waste;

[0071] Step S22: For each waste category, use a large language model to generate specific descriptive text with multiple semantic dimensions according to each semantic ontology;

[0072] Step S23: Encode the specific descriptive text of each semantic dimension into text description features through a text encoder to form a structured semantic knowledge base corresponding to the category, and calculate the dimension prototype vector for each semantic dimension. The dimension prototype vector is used to represent the semantic centroid of the semantic dimension in the feature space.

[0073] Example 5: This example is based on the above examples. The step of decomposing the text description features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism includes:

[0074] Step S31: Construct a dimensional decomposition semantic refinement module, which includes multi-dimensional expert groups corresponding to each semantic dimension, and each dimension expert group contains several expert query matrices;

[0075] Step S32: Set up the first-level routing network to calculate the dimensional importance weight vector of each semantic dimension based on the multi-scale visual summarization vector: ;

[0076] In the formula, This refers to the multi-scale visual summarization vector. This is the first-level routing weight matrix, used to map visual features to a dimensionality importance space. This is the first-level routing bias vector. Represents the activation function. Represents the dimension importance weight vector;

[0077] Step S33: Calculate the routing score for each expert based on textual description features within each semantic dimension using the second-level routing network: ;

[0078] In the formula, Represents textual description features. This represents a linear projection function specific to dimension a. The Gaussian noise term representing input dependency is used to prevent routing network collapse and promote diversity and specialization in expert learning. The expert routing score matrix is ​​represented by dimension a;

[0079] Step S34: Based on the routing score, select some experts for each semantic dimension, and use the expert query matrix corresponding to the selected experts to decompose the text description features into fine-grained semantic clues corresponding to the dimensions and experts.

[0080] Example 6: This example is based on the above examples. The step of generating sample-level semantic priors includes:

[0081] Step S35: Calculate multi-head cross attention using the fine-grained semantic clues as the query and the local texture feature matrix as the key and value;

[0082] Step S36: Incorporate the dimensional importance weights into the cross-attention calculation to adaptively enhance the attention response of specific semantic dimensions based on the visual characteristics of the current sample;

[0083] Step S37: Weight and fuse the expert outputs in each semantic dimension according to the expert routing score to obtain the fused representation of each dimension;

[0084] Step S38: The fusion representations of each dimension are weighted and fused according to the importance weight of the dimensions to obtain a category-level semantic representation;

[0085] Step S39: Aggregate the category-level semantic representations of all categories to obtain the sample-level semantic prior.

[0086] Example 7: Based on the above examples, the step of retrieving matching multi-perspective text features from the structured semantic knowledge base includes:

[0087] Step S41: Align the sample-level semantic priors in the feature space to obtain adapted prior features;

[0088] Step S42: Calculate the dimension-aware comprehensive similarity measure based on the dimension importance weights, the similarity between the adapted prior features and the prototype vectors of each dimension, and the similarity between the text description features and the prototype vectors of each dimension.

[0089] Step S43: Based on the comprehensive similarity metric, select several text description features with the highest similarity from the structured semantic knowledge base of the corresponding category to form a retrieval set;

[0090] Step S44: Randomly shuffle the retrieved set and store it in a queue so that the random optimization strategy can use it in a random order.

[0091] Example 8: This example is based on the above examples. The step of generating auxiliary semantic priors using the retrieved multi-view text features and training the classifier by jointly using the main prior and auxiliary priors through a random optimization strategy includes:

[0092] Step S45: In each training iteration, pop a text description feature from the queue;

[0093] The pop-up text description features are decomposed into auxiliary fine-grained semantic cues through the decompositional semantic refinement mechanism, and auxiliary semantic priors are generated through visual-text cross-attention interaction.

[0094] Step S46: Input the sample-level semantic prior and the auxiliary semantic prior into the shared classifier to obtain the main classification result and the auxiliary classification result, and then fuse the two to obtain the final classification result;

[0095] Step S47: Calculate the classification loss based on the final classification result and the true label, and update the model parameters with the classification loss as the optimization objective.

[0096] Example 9: Based on the above examples, this example provides a waste sorting system based on intelligent analysis, including a visual feature extraction module, a structured semantic knowledge base construction module, a semantic prior generation module, a classifier training module, and a waste sorting module.

[0097] The visual feature extraction module acquires images of the waste items to be classified and extracts their visual features.

[0098] The structured semantic knowledge base construction module constructs a structured semantic knowledge base for the field of waste classification. The semantic knowledge base contains text description features with multiple preset semantic dimensions.

[0099] The semantic prior generation module uses a decompositional semantic refinement mechanism to decompose textual descriptive features into multiple fine-grained semantic cues, and generates sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues.

[0100] The classifier training module, based on the sample-level semantic prior, retrieves matching multi-perspective text features from the structured semantic knowledge base, generates auxiliary semantic priors using the retrieved multi-perspective text features, and trains the classifier model by combining the main prior and auxiliary priors through a random optimization strategy.

[0101] The waste sorting module inputs images of waste to be sorted into the trained model to obtain waste sorting results.

[0102] Example 10: Based on the above examples, this example uses knowledge distillation technology to use the fully trained model as the teacher model and transfer its semantic enhancement capabilities to a lightweight pure vision student model.

[0103] The student model only needs to receive visual input during the inference phase to output classification results, thus adapting to the resource constraints of edge computing devices.

[0104] Example 11: Based on the above examples, the specific settings of the semantic ontology in step S21 are as follows:

[0105] Define a four-dimensional semantic ontology for garbage classification:

[0106] Dimension 1, Material Properties:

[0107] Describe the material composition of the item, for example: "Transparent PET plastic with a smooth and glossy surface";

[0108] Dimension Two, Shape and Structure:

[0109] Describe the geometric shape and structural features of an object, for example: "cylindrical bottle body, narrow neck, threaded mouth";

[0110] Dimension 3, Usage Category:

[0111] Describe the original function and type of contents of the item, for example: "carbonated beverage packaging container";

[0112] Dimension Four: Handling Methods;

[0113] Describe the recyclability of the item and precautions for handling it. Example: "Non-recyclable composite material, made of multiple layers of paper, plastic and aluminum."

[0114] 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, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations 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.

[0116] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A waste sorting method based on intelligent analysis, characterized in that: Specifically, the following steps are included: Step S1: Obtain images of the waste items to be classified and extract their visual features; Step S2: Construct a structured semantic knowledge base for the field of waste classification, wherein the semantic knowledge base contains textual description features with multiple preset semantic dimensions; Step S3: Decompose the text description features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism, and generate sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues. Step S4: Based on the sample-level semantic prior, retrieve matching multi-perspective text features from the structured semantic knowledge base, generate auxiliary semantic prior using the retrieved multi-perspective text features, and train the classifier model by combining the main prior and auxiliary prior through a random optimization strategy. Step S5: Input the garbage image to be classified into the trained model to obtain the garbage classification result.

2. The waste sorting method based on intelligent analysis according to claim 1, characterized in that: The steps for extracting its visual features include: Step S11: Use parallel local texture encoders and global shape encoders to extract the local texture feature matrix and global shape feature vector of the garbage image, respectively; Step S12: Calculate the fusion weights through an adaptive gating network, and perform weighted fusion of the pooled representation of local texture features and the global shape feature vector according to the fusion weights to generate a multi-scale visual summary vector, i.e., visual features.

3. The waste sorting method based on intelligent analysis according to claim 2, characterized in that: The adaptive gating network calculates the fusion weights. The formula is: ； In the formula, For local texture feature matrix, This is an image-level local representation obtained by average pooling along the local region dimension; This is the global shape feature vector after linear projection alignment; This represents a vector concatenation operation; It is a learnable gated weight matrix; This is a gated bias term; Use the Sigmoid activation function; For weighting; The multi-scale visual summarization vector The calculation formula is: 。 4. The waste sorting method based on intelligent analysis according to claim 3, characterized in that: The steps for constructing a structured semantic knowledge base for the field of waste sorting include: Step S21: Define a semantic ontology for waste classification that includes the dimensions of material properties, shape and structure, usage category, and disposal method of waste; Step S22: For each waste category, use a large language model to generate specific descriptive text with multiple semantic dimensions according to each semantic ontology; Step S23: Encode the specific descriptive text of each semantic dimension into text description features through a text encoder to form a structured semantic knowledge base corresponding to the category, and calculate the dimension prototype vector for each semantic dimension. The dimension prototype vector is used to represent the semantic centroid of the semantic dimension in the feature space.

5. The waste sorting method based on intelligent analysis according to claim 4, characterized in that: The step of decomposing textual descriptive features into multiple fine-grained semantic cues using a decompositional semantic refinement mechanism includes: Step S31: Construct a dimensional decomposition semantic refinement module, which includes multi-dimensional expert groups corresponding to each semantic dimension, and each dimension expert group contains several expert query matrices; Step S32: Set up the first-level routing network to calculate the dimensional importance weight vector of each semantic dimension based on the multi-scale visual summarization vector: ； In the formula, This refers to the multi-scale visual summarization vector. This is the first-level routing weight matrix, used to map visual features to a dimensionality importance space. This is the first-level routing bias vector. Represents the activation function. Represents the dimension importance weight vector; Step S33: Calculate the routing score for each expert based on textual description features within each semantic dimension using the second-level routing network: ； In the formula, Represents textual description features. This represents a linear projection function specific to dimension a. The Gaussian noise term representing input dependency is used to prevent routing network collapse and promote diversity and specialization in expert learning. The expert routing score matrix is ​​represented by dimension a; Step S34: Based on the routing score, select some experts for each semantic dimension, and use the expert query matrix corresponding to the selected experts to decompose the text description features into fine-grained semantic clues corresponding to the dimensions and experts.

6. The waste sorting method based on intelligent analysis according to claim 5, characterized in that: The steps for generating sample-level semantic priors include: Step S35: Calculate multi-head cross attention using the fine-grained semantic clues as the query and the local texture feature matrix as the key and value; Step S36: Incorporate the dimensional importance weights into the cross-attention calculation to adaptively enhance the attention response of specific semantic dimensions based on the visual characteristics of the current sample; Step S37: Weight and fuse the expert outputs in each semantic dimension according to the expert routing score to obtain the fused representation of each dimension; Step S38: The fusion representations of each dimension are weighted and fused according to the importance weight of the dimensions to obtain a category-level semantic representation; Step S39: Aggregate the category-level semantic representations of all categories to obtain the sample-level semantic prior.

7. A waste sorting method based on intelligent analysis according to claim 6, characterized in that: The steps for retrieving matching multi-perspective text features from a structured semantic knowledge base include: Step S41: Align the sample-level semantic priors in the feature space to obtain adapted prior features; Step S42: Calculate the dimension-aware comprehensive similarity measure based on the dimension importance weights, the similarity between the adapted prior features and the prototype vectors of each dimension, and the similarity between the text description features and the prototype vectors of each dimension. Step S43: Based on the comprehensive similarity metric, select several text description features with the highest similarity from the structured semantic knowledge base of the corresponding category to form a retrieval set; Step S44: Randomly shuffle the retrieved set and store it in a queue so that the random optimization strategy can use it in a random order.

8. A waste sorting method based on intelligent analysis according to claim 7, characterized in that: The steps of generating auxiliary semantic priors using retrieved multi-view text features and training the classifier by jointly using the main prior and auxiliary priors through a random optimization strategy include: Step S45: In each training iteration, pop a text description feature from the queue; The pop-up text description features are decomposed into auxiliary fine-grained semantic cues through the decompositional semantic refinement mechanism, and auxiliary semantic priors are generated through visual-text cross-attention interaction. Step S46: Input the sample-level semantic prior and the auxiliary semantic prior into the shared classifier to obtain the main classification result and the auxiliary classification result, and then fuse the two to obtain the final classification result; Step S47: Calculate the classification loss based on the final classification result and the true label, and update the model parameters with the classification loss as the optimization objective.

9. A waste sorting system based on intelligent analysis, used to implement the waste sorting method based on intelligent analysis as described in any one of claims 1 to 8, characterized in that: It includes a visual feature extraction module, a structured semantic knowledge base construction module, a semantic prior generation module, a classifier training module, and a garbage classification module; The visual feature extraction module acquires images of the waste items to be classified and extracts their visual features. The structured semantic knowledge base construction module constructs a structured semantic knowledge base for the field of waste classification. The semantic knowledge base contains text description features with multiple preset semantic dimensions. The semantic prior generation module uses a decompositional semantic refinement mechanism to decompose textual descriptive features into multiple fine-grained semantic cues, and generates sample-level semantic priors through visual-text cross-attention interaction. The sample-level semantic priors integrate visual features and fine-grained semantic cues. The classifier training module, based on the sample-level semantic prior, retrieves matching multi-perspective text features from the structured semantic knowledge base, generates auxiliary semantic priors using the retrieved multi-perspective text features, and trains the classifier model by combining the main prior and auxiliary priors through a random optimization strategy. The waste sorting module inputs images of waste to be sorted into the trained model to obtain waste sorting results.

Images