A feature subset optimization method and system for garbage image classification

Abstract

The application relates to the technical field of data processing, and discloses a feature subset optimization method and system for garbage image classification, which comprises the following steps: performing feature extraction and spatial activation feature weighting on an image dataset to obtain a weighted visual feature full set; selecting initial features from the weighted visual feature full set and determining an effective candidate set; calculating single feature information contribution, joint information contribution and background co-response redundancy of each subset in the effective candidate set to screen each subset, obtaining multiple non-dominated subsets to construct a non-dominated subset set; selecting a subset with the maximum subset contribution degree from the non-dominated subset set as the initial feature of the next iteration process and performing iteration; and outputting a final optimal optimization feature subset when the iteration meets a termination condition; the application solves the problems of poor generalization ability and robustness and low calculation efficiency of existing feature screening methods.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a feature subset optimization method and system for garbage image classification. Background Technology

[0002] With the rapid development of artificial intelligence and machine learning technologies, image data modeling methods based on high-dimensional feature representations have been widely applied in image classification, object recognition, pattern recognition, and intelligent decision-making. In these applications, convolutional neural networks or other image feature extraction methods are typically used to map the original image data into high-dimensional feature vectors, and model training and classification prediction are then performed based on the obtained feature representations. The quality of the feature representation largely determines the model performance; therefore, feature selection is a crucial step in image data processing, significantly contributing to improved classification accuracy and reduced computational complexity.

[0003] In waste image classification tasks, different waste categories often exhibit strong visual similarities. For example, plastic, paper, and metal packaging show high overlap in color texture, edge structure, and local morphology. Simultaneously, due to the complex environment in which waste images are captured, images typically contain background interference, occlusion, lighting variations, and noise from irrelevant areas. This leads to a large number of visual feature channels in convolutional features that produce similar responses to the same background regions, repetitive texture regions, or local visual patterns. While these visual feature channels have high response values ​​in the feature space, their corresponding information contributions overlap significantly, easily resulting in redundant visual representations and weakening the classification model's ability to discriminate key waste target regions. Directly using all convolutional features for classification not only increases the model's computational complexity but may also lead to decreased stability of classification results due to the repeated superposition of redundant visual patterns. Therefore, it is necessary to effectively filter convolutional visual features to address the feature redundancy problem caused by complex backgrounds and repetitive visual patterns in waste images, obtaining more compact visual feature representations with class discriminative capabilities.

[0004] Currently, existing feature selection methods mainly include filtering methods, wrapping methods, and embedded methods. Filtering methods typically evaluate features independently based on statistical metrics such as correlation coefficients, variance, or mutual information. However, these methods often ignore the combination relationships between features and the overall distribution structure, making it difficult to effectively identify co-response visual features caused by background regions or repetitive texture regions in junk images. Wrapping methods combine specific learning models and use model performance as the evaluation criterion, but their computational complexity is high and they depend on specific model structures, easily leading to decreased search efficiency in high-dimensional junk image convolutional feature scenarios. Embedded methods complete feature selection during model training, but their results are usually limited by the model form and lack stable generalization ability for visual redundancy patterns in complex junk scenes.

[0005] Furthermore, in existing technologies related to waste image classification, the technical solution with publication number CN119478531A proposes an automated image recognition waste classification process based on AI computing. This method integrates image data, weight data, and near-infrared spectral data, combined with multi-scale target detection, attention mechanisms, and probability density distribution modeling, to achieve automatic identification and classification of waste targets. This type of method focuses on waste target detection and multi-modal classification process construction, emphasizing the fusion of multi-source sensor information and improved detection accuracy. However, its feature processing mainly relies on a fixed convolutional network structure and attention enhancement mechanism, lacking specialized modeling of the redundancy relationships within high-dimensional convolutional features and the perturbation characteristics of category distribution. This makes it difficult to effectively filter repetitive visual patterns, background interference features, and low-contribution channels at the feature subset level.

[0006] Furthermore, the technical solution published in CN121425686A proposes an intelligent sorting trash can based on image recognition. It uses a convolutional neural network to extract features and classify trash images, combining a saliency enhancement module, an attention module, and an automatic sorting mechanism to achieve trash category identification and intelligent disposal control. This type of method mainly focuses on trash sorting equipment and image recognition hardware systems, emphasizing the design of the sorting execution structure and visual recognition process. However, its feature representation still primarily relies on the high-dimensional feature results directly output by the convolutional network, lacking a feature redundancy suppression mechanism for background noise, visual co-response features, and overlapping category regions in trash images. It also fails to establish a feature subset optimization process based on feature distribution perturbation and joint information contribution. Therefore, in complex trash scenarios, it is prone to problems such as repetitive feature representation and insufficient classification stability.

[0007] It is evident that existing technologies still lack a feature subset optimization method that can effectively model complex backgrounds, repetitive visual patterns, and high-dimensional convolutional feature redundancy in the context of garbage image classification. This results in problems such as high feature redundancy, insufficient classification stability, and low search efficiency in the garbage image classification process. Summary of the Invention

[0008] This invention provides a feature subset optimization method and system for garbage image classification, in order to solve the problems of redundant convolutional features, decreased class discrimination ability, and insufficient classification stability caused by background interference, complex textures, and repetitive visual responses during garbage image classification.

[0009] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a feature subset optimization method for garbage image classification, comprising: Image datasets are constructed by acquiring image data of garbage images. Convolutional feature extraction and spatial activation feature weighting are performed on the garbage images to obtain a complete set of convolutional visual features that represent the color texture, edge structure and local morphological information of the garbage region. The complete set of convolutional visual features is pre-weighted based on the spatial activation concentration of each convolutional visual feature channel to obtain a weighted complete set of visual features. Set the maximum number of iterations, randomly select a sub-feature from the full set of weighted visual features as the initial feature, and perform a single-step transition operation on the initial feature to construct a candidate set. Then, determine the effective candidate set from the candidate set through the constructed feature information contribution evaluation function. The single feature information contribution and joint information contribution of each subset in the effective candidate set are calculated based on the feature information contribution evaluation function, and the background co-response redundancy of each subset is calculated based on the degree of deviation between the single feature information contribution and the joint information contribution. The number of features and contribution of each subset are determined. Based on the contribution, background co-response redundancy and the number of features, each subset is filtered to obtain multiple non-dominated subsets to construct a set of non-dominated subsets. Construct an optimal feature subset, select the optimal feature subset from the non-dominated subset set as the initial feature of the next iteration process and update it to the optimal feature subset, and iterate based on the initial feature of the next iteration process to obtain the non-dominated subset set; In each iteration, the best feature subset of the current iteration is compared with the optimal optimized feature subset. The optimal optimized feature subset is updated based on the comparison result. When the iteration meets the termination condition, the updated best feature subset is output as the final optimal optimized feature subset to classify garbage images.

[0010] Secondly, the present invention also provides a feature subset optimization system for garbage image classification, including a processor and a memory; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the steps of the method described in the first aspect.

[0011] Beneficial effects: This invention provides a feature subset optimization method for garbage image classification. By constructing a feature information contribution evaluation function, it directly measures the perturbation strength of the feature subset conditional distribution relative to the uninformed prior distribution. This quantifies the influence of different convolutional visual feature channels on the garbage category distribution from the fundamental level of data distribution, and can distinguish the visual response differences between real garbage target areas and background areas. It solves the problem of existing filtering methods neglecting the combination relationship between features and the overall distribution structure due to reliance on univariate statistical indicators. The method defines background co-response redundancy by calculating the additivity deviation between single feature information contribution and joint information contribution, and simultaneously incorporates subset contribution, background co-response redundancy, and feature quantity into the dominance filtering rule to select non-dominated subsets. While improving feature discrimination ability, it also identifies the co-response phenomenon of multiple convolutional feature channels to background areas, lighting textures, and repetitive structures in garbage images by calculating the deviation between single feature information contribution and joint information contribution, and suppresses the simultaneous selection of such repetitive visual response features into feature subsets. This balances feature dimension simplification and model structure compactness, overcoming the limitations of existing wrapper methods. This method avoids the drawbacks of high computational complexity and over-reliance on specific model structures. By setting a single-step transition operation to add or delete features from the current feature subset to generate a candidate set, and combining the memory mechanism of the visited state set with gain filtering to screen effective candidates, it achieves efficient search in a large feature combination space, avoiding redundant calculations and ineffective expansion. It significantly improves the computational efficiency and convergence stability of feature selection while ensuring search accuracy. By dynamically maintaining the globally optimal feature subset in each iteration and outputting the final optimal feature subset when the termination condition is met, this method can complete feature screening without relying on a specific classification model structure. The output final optimal feature subset improves the generalization ability and robustness of the trained classification model. It is particularly suitable for garbage image classification scenarios with strong background interference, large variation in garbage target size, high visual similarity of categories, and obvious local pollution occlusion. By pre-weighting the spatial activation concentration, the geometric spatial distribution information of the convolutional feature map is incorporated into the feature screening entry point, completing the initial screening of garbage target-background at the physical geometry level before the information domain evaluation, realizing dual-domain collaborative feature screening of the geometric and information domains. Attached Figure Description

[0012] Figure 1This is a flowchart of a preferred embodiment of the feature subset optimization method for garbage image classification according to the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example 1 Please see Figure 1 This application provides a feature subset optimization method for garbage image classification, including: Image datasets are constructed by acquiring image data of garbage images. Convolutional feature extraction and spatial activation feature weighting are performed on the garbage images to obtain a complete set of convolutional visual features that represent the color texture, edge structure and local morphological information of the garbage region. The complete set of convolutional visual features is pre-weighted based on the spatial activation concentration of each convolutional visual feature channel to obtain a weighted complete set of visual features. Set the maximum number of iterations, randomly select a sub-feature from the full set of weighted visual features as the initial feature, and perform a single-step transition operation on the initial feature to construct a candidate set. Then, determine the effective candidate set from the candidate set through the constructed feature information contribution evaluation function. The single feature information contribution and joint information contribution of each subset in the effective candidate set are calculated based on the feature information contribution evaluation function, and the background co-response redundancy of each subset is calculated based on the deviation between the single feature information contribution and the joint information contribution, so as to suppress the repeated responses of multiple feature channels to background regions, repetitive textures and irrelevant visual regions in garbage images. The number of features and contribution of each subset are determined. Based on the contribution, background co-response redundancy and the number of features, each subset is filtered to obtain multiple non-dominated subsets to construct a set of non-dominated subsets. Construct an optimal feature subset, select the optimal feature subset from the non-dominated subset set as the initial feature of the next iteration process and update it to the optimal feature subset, and iterate based on the initial feature of the next iteration process to obtain the non-dominated subset set; In each iteration, the best feature subset of the current iteration is compared with the optimal optimized feature subset. The optimal optimized feature subset is updated based on the comparison result. When the iteration meets the termination condition, the updated best feature subset is output as the final optimal optimized feature subset to classify garbage images.

[0015] In the above embodiments, by acquiring a dataset of garbage images and performing feature extraction and spatial activation feature weighting, a weighted visual feature set is obtained, providing a unified data foundation for subsequent feature selection. Setting a maximum number of iterations and randomly selecting initial features from the weighted visual feature set avoids the influence of manually preset biases on search results and ensures diversity of search starting points. A candidate set is constructed through single-step transition operations, performing only one addition or deletion operation on the current subset each time, ensuring fine-grained search and avoiding skipping potential optimal solutions with excessively large steps. A feature information contribution evaluation function is constructed to effectively screen the candidate set, retaining only candidate subsets with substantial improvements in information contribution, avoiding the waste of computational resources from ineffective expansion. Based on this evaluation function, the single feature information contribution and joint information contribution of each subset are calculated, and the background co-response redundancy is calculated accordingly. This quantifies the substitutability between features from the perspective of information overlap, providing a key basis for subsequent multi-objective evaluation. The selection system incorporates subset contribution, background co-response redundancy, and feature quantity simultaneously. Candidate subsets are divided into dominant and non-dominated subsets based on dominance relationships, achieving a unified trade-off between the three mutually constraining objectives of information content, compactness, and low redundancy. The subset with the highest contribution from the non-dominated subset is selected as the initial feature for the next iteration. In each iteration, the current optimal feature is compared with the global optimal feature to ensure the search process continuously evolves towards higher information contribution. When the iteration meets the termination condition, the final optimal feature subset is output, ensuring both a clear convergence guarantee and no overlooking of the global optimal solution.

[0016] Preferably, image datasets are constructed by acquiring image data of garbage images, and convolutional feature extraction is performed on the garbage images to obtain a complete set of convolutional visual features representing the color texture, edge structure, and local morphological information of the garbage region, including: By acquiring garbage images and performing a unified modeling on them, a garbage image dataset is obtained, which satisfies the following relationship: ; In the formula, Represents an image dataset. Indicates the first A sample of garbage images, This indicates the corresponding waste category label. Indicates the total number of samples; The image feature extraction model extracts the mapped data from the image dataset. The feature representation of the 1D feature vector yields the corresponding complete set of convolutional visual features, where the complete set of convolutional visual features satisfies the following relationship: ; In the formula, Represents the complete set of visual features from convolution. Represents the first feature map in the convolutional feature map. The response features corresponding to each visual feature channel Indicates the dimension of image features.

[0017] In the above embodiments, a unified modeling of garbage images yields an image dataset that satisfies a relational expression. This transforms the original unstructured image data into structured sample-label pairs, enabling subsequent feature extraction and subset evaluation to be processed in batches based on a standardized data format. A pre-defined image feature extraction model maps each image to a 3D feature vector, forming a convolutional visual feature set. This transforms the original pixel-level image representation into a high-dimensional feature representation with semantic discriminative capabilities. Each dimension's feature response corresponds to a different visual pattern, providing high-dimensional visual response units for subsequent feature subset selection. The dimension of the convolutional visual feature set serves as the upper bound of the search space, defining the scale boundary of the feature subset search problem.

[0018] In this embodiment, a garbage image is input into a preset convolutional neural network. The edge textures, color regions, and local structures in the image are encoded layer by layer through convolutional and pooling layers. Visual feature representations are extracted from high-level convolutional feature maps. The total number of channels in the convolutional feature map is set to... Then, the response feature corresponding to each channel is defined as a candidate visual feature, and thus a complete set of visual convolutional visual features is constructed. .

[0019] The entire set of convolutional visual features is pre-weighted based on the spatial activation concentration of each convolutional visual feature channel to obtain a weighted set of visual features, including: Obtain the output of the last convolutional module of the convolutional neural network 3D feature map, where These represent the spatial height and width of the feature map, respectively. Number of channels; For each channel Calculate the feature map of this channel in The activation concentration in space, which satisfies the following relationship: ; ; ; In the formula, Indicates the first The spatial location of each channel Activation value at that location, This represents the spatial activation mean of the channel. This represents the spatial activation variance of the channel. A preset small constant to prevent division by zero; Channel weights are generated based on the activation concentration of each channel, and the feature values ​​of each channel after global average pooling are weighted. The channel weights and weighted feature values ​​satisfy the following relationships: ; ; In the formula, Represents the normalized i-th Each channel weight, Indicates the weighted number of... Each channel feature value, Represents the first digit after the original global average pooling. Each channel feature value, This represents the preset weight scaling factor. ; All weighted channel feature values ​​constitute the complete set of weighted visual features.

[0020] In the above embodiments, the physical geometric characteristics of garbage target regions in garbage images—characterized by locally concentrated high activation responses in the convolutional feature map, while background regions such as conveyor belts or container walls exhibit widely dispersed activation responses—are utilized. The geometric focusing degree of each channel on the garbage target is quantified by calculating the ratio of its spatial activation variance to its mean. Channels with high concentration have activation responses concentrated in a small area, corresponding to garbage target features; channels with low concentration have activation responses evenly distributed across the entire space, corresponding to background scattering features. Channel weights are generated through normalization and then... The original feature values ​​are weighted to ensure that even the channels with the lowest concentration are not completely zeroed out—because these channels may still carry information with marginal discriminative value regarding the distribution of waste categories at the information theory level, and geometric domain features alone are insufficient to completely exclude them. Weight scaling factor. Controlling the pre-weighting strength, when The pre-weighted state degenerates into an unweighted state, facilitating ablation experiments to verify the effectiveness of the pre-weighting technique. Spatial activation concentration pre-weighting and subsequent information domain feature evaluation based on KL divergence operate in the geometric domain and information domain, respectively. They are mutually complementary and cannot be derived or substituted for each other. After completing geometric domain pre-weighting at the feature screening entry point, the initial feature space entering the subsequent iterative evaluation process has preliminarily distinguished between the garbage target channel and the background scattering channel, effectively compressing the search space scale.

[0021] Preferably, a sub-feature is randomly selected from the entire set of weighted visual features as the initial feature, and a candidate set is constructed by performing a single-step transition operation on the initial feature, including: Randomly select a sub-feature from the weighted set of visual features as the initial feature, and define the initial feature to satisfy the following relationship: ; In the formula, Indicates initial features, This represents a feature index randomly selected from the entire set of weighted visual features. Indicates the maximum number of iterations; The single-step transition operation includes addition and deletion operations. By performing the addition and deletion operations in the single-step transition operation on the initial features, a neighborhood candidate set of the current state is generated. The add operation includes: Each sub-feature in the weighted visual feature set is used as an initial feature. For each initial feature, if the corresponding initial feature does not exist in the current feature subset state... When the initial feature is added to the current feature subset state, a neighborhood set is generated. The neighborhood set satisfies the following relationship: ; In the formula, Indicates the first The neighborhood set added in the next iteration; Deletion operations include: For each initial feature, the corresponding initial feature is removed from the current feature subset state, generating a deletion neighborhood set. The deletion neighborhood set satisfies the following relation: ; In the formula, Indicates the first The next iteration corresponds to the deletion of the neighborhood set. Represents the first element in the weighted visual feature set. Individual characteristics; A candidate set is constructed based on adding to and deleting from the neighborhood set. The candidate set satisfies the following relationship: ; In the formula, Indicates the first The candidate set generated in each iteration. Each iteration only expands the first-order neighborhood of the current state and only evaluates and filters the neighborhood candidates, rather than performing a global enumeration of all feature combinations.

[0022] In the above embodiments, by randomly selecting a feature from the entire set of weighted visual features as the initial feature and using a single-element subset as the starting point for the search, the search process can begin exploring from any feature dimension, avoiding subjective bias towards specific features. A single-step transition mechanism including addition and deletion operations is constructed. The addition operation generates an addition subset for features not in the current subset, and the deletion operation generates a deletion subset by removing the corresponding feature from the current subset. Both operations only change the state of one feature, forming a first-order neighborhood search structure. The candidate set consists of the union of the addition and deletion subsets, ensuring that the search process can both expand the discriminative capability by adding new features and simplify the structure by deleting redundant or invalid features, achieving bidirectional reversible exploration within the feature subset space.

[0023] Preferably, the effective candidate set is determined from the candidate set by constructing a feature information contribution evaluation function, including: For any subset in the candidate set, the similarity of the subset in the sample space is calculated using a Gaussian kernel function, where the Gaussian kernel function satisfies the following relationship: ; In the formula, Both indicate that the sample is in the feature subset The projection vector on, Denotes the Euclidean norm. This represents the kernel bandwidth parameter of the corresponding subspace, and its candidate set is: ,in Representing a feature subset The standard deviation of the pairwise Euclidean distance of the samples is adaptively determined based on the statistical characteristics of the sample distance distribution in the feature subspace; This invention models the category distribution changes induced by convolutional features in garbage images, characterizing the contribution of different visual feature channels to garbage category discrimination. It also combines feature redundancy constraints to suppress the influence of background noise and repetitive visual patterns, thereby improving the expressive power of key visual features in garbage image classification. In this embodiment, a label distribution for any subset is constructed based on a Gaussian kernel function, where the label distribution satisfies the following relationship: ; In the formula, Indicating in the feature subset Conditional Labels The probability estimate, This is an indicator function; it takes the value 1 when the indicator function condition is true, and 0 otherwise. Indicates the first The feature labels corresponding to each sample Indicates sample and In feature subset projection and Spatial similarity; Construct the label prior distribution under featureless conditions, where the label prior distribution under featureless conditions satisfies the following relationship: ; In the formula, This represents the label distribution without using any feature information. Indicates the first The feature labels corresponding to each sample; A feature information contribution evaluation function is constructed based on the label distribution of arbitrary subsets and the prior label distribution under featureless conditions. The feature information contribution evaluation function satisfies the following relationship: ; ; In the formula, Indicates the Kullback–Leibler divergence. Representing a feature subset Compared to the intensity of distribution perturbation introduced by the informationless state, the distribution of different waste categories in the visual feature space varies significantly. Therefore, when key visual features are preserved, their corresponding label distributions will deviate more significantly from the featureless state; while redundant or noisy features usually only cause local distribution fluctuations. Unlike symmetric information measures such as class mutual information, KL divergence, as an asymmetric divergence measure, measures the degree of one-way deviation of the category distribution induced by feature subsets from the informationless prior state, rather than the amount of bidirectional information sharing between features and labels. When two visual feature channels encode highly overlapping waste category discrimination information, both can obtain high class mutual information values, but their corresponding KL divergence evaluation values ​​will show significant differences due to feature information redundancy—because the second feature hardly contributes any additional deviation to the label distribution reshaped by the first feature. Thus, feature discrimination evaluation and feature redundancy relationship identification can be achieved simultaneously within a unified information theory metric space without the need to introduce auxiliary metric standards. Therefore, this invention uses distribution difference as the evaluation basis for visual feature information contribution. The first feature information contribution evaluation function value of the initial features is calculated based on the feature information contribution evaluation function. The evaluation function value of the contribution of the second feature information of all subsets in the candidate set. ; Filter all contribution evaluation function values ​​of the second feature information The value of the contribution evaluation function of the first feature information is greater than A new set of valid candidates is constructed from a subset of the subsets, wherein the construction of the set of valid candidates satisfies the following relation: ; In the formula, Denotes the set of valid candidates. Denotes a subset of the candidate set. Denotes the candidate set, This indicates that the set has been visited.

[0024] In the above embodiments, the similarity of subsets in the sample space is calculated using a Gaussian kernel function. Weighted information from local neighborhoods is used to characterize the distribution relationship between visual features of garbage images. The introduction of the kernel bandwidth parameter allows the similarity decay rate to be adaptively adjusted according to the statistical characteristics of the subspace, avoiding the measurement distortion problem caused by a fixed scale in different dimensional subspaces. A label distribution for arbitrary subsets is constructed based on the Gaussian kernel function. By weighting and globally averaging the local neighborhoods of each sample, the visual proximity relationship in the convolutional feature space of garbage images is transformed into a garbage category probability distribution, realizing the transformation from "feature similarity" to "category discriminability." A prior label distribution under featureless conditions is constructed as a reference benchmark, reflecting the original cognitive level when relying solely on the frequency of samples from each category, providing a unified reference system for subsequent distribution difference measurement. A feature information contribution evaluation function is constructed based on KL divergence. The information contribution of the feature subset is measured by quantifying the deviation of the garbage category distribution induced by the visual feature subset from the prior distribution. A larger evaluation value indicates a more significant improvement in category cognition after using the subset. By comparing the evaluation values ​​of candidate subsets with the evaluation values ​​of the current subset, only visual feature subsets that can enhance the ability to distinguish the distribution of waste categories are retained to construct a new effective candidate set. In addition, the visited set is used for deduplication filtering. This dual screening mechanism ensures the directionality of each search step and avoids repeated exploration of historical paths.

[0025] Preferably, the single feature information contribution and joint information contribution of each subset in the effective candidate set are calculated based on the feature information contribution evaluation function, including: Calculate a single feature subset based on the feature information contribution evaluation function. The evaluation value is used to obtain the single feature information contribution of the corresponding feature. ,and, ; Calculate the binary feature subset { based on the feature information contribution evaluation function. The evaluation value is used to obtain the joint information contribution of the corresponding feature pairs. ,and, .

[0026] In the above embodiments, the single-feature information contribution of each feature in the effective candidate set is calculated using a feature information contribution evaluation function. This measures the ability of a single visual feature channel to independently classify garbage target regions. A larger information contribution value indicates a stronger reshaping effect on the category distribution when the feature is used alone. For subsets containing two or more features, the joint information contribution of any two features is further calculated to measure the comprehensive discrimination effect of the features when used collaboratively. The joint information contribution and the single-feature information contribution are calculated using the same evaluation function, ensuring consistency in the measurement standard and allowing for direct numerical comparison. This lays the computational foundation for quantifying redundancy relationships through additivity deviation.

[0027] Preferably, the background co-response redundancy of each subset is calculated based on the deviation between the contribution of single feature information and the contribution of joint information, including: The background co-response redundancy of all subsets of the effective candidate set is calculated based on the difference between the contribution of single feature information and the contribution of joint information. The calculation of the background co-response redundancy satisfies the following relationship: ; In the formula, Represents any feature subset Background co-response redundancy, , Both represent contributions from single feature information. This represents the joint feature information contribution of the two features.

[0028] In garbage images, the background area accounts for a larger proportion than the target area, and the convolutional channels are prone to collective activation of the background texture, resulting in multiple visual channels jointly encoding the background pattern. This invention identifies such co-response channels by joint information gain bias and suppresses background co-response features from entering the final subset during the feature selection process.

[0029] It is worth noting that traditional redundancy measurement methods based on mutual information or correlation coefficients between features can only identify visual feature channels with similar activation response patterns, but cannot identify feature pairs with different visual response patterns that are synchronously activated in the background region, resulting in a lack of additivity in their joint discriminative contributions. This invention defines background co-response redundancy by using the additivity deviation of the sum of joint information contributions and single feature information contributions to elevate redundancy identification from response similarity detection to discriminative information additivity detection. Within a unified KL divergence evaluation framework, it can identify background co-response redundant feature pairs that cannot be detected solely by feature response distribution similarity.

[0030] In the above embodiments, the background co-response redundancy of each subset is calculated based on the difference between single feature information contribution and joint information contribution, directly quantifying the degree of information overlap between features. In convolutional features of garbage images, multiple convolutional visual channels may simultaneously respond to similar pollution texture regions, background structures, packaging color patterns, or edge contour regions in the garbage image, resulting in repeated responses to the same garbage visual patterns. If such co-response features are retained simultaneously, the classification model may overemphasize the garbage background region, thereby weakening its ability to classify the real garbage target region and increasing the risk of misclassification between different garbage categories. The calculation method of background co-response redundancy embodies the idea of ​​marginal contribution analysis: if the information provided by two features does not overlap at all, the joint contribution is equal to the sum of their individual contributions, and the value of the redundancy term is zero; if the information provided by two features is highly repetitive, the increment of their joint information contribution relative to the sum of their individual contributions will weaken, thereby increasing the value of the corresponding redundancy term accordingly. By summing the redundant terms of all non-repeating feature pairs, the background co-response redundancy at the entire subset level is obtained. This allows the background co-response redundancy measure to accumulate naturally as the subset size increases, effectively suppressing visual channels that produce repeated responses to the same garbage target region or background region from being repeatedly selected into the same feature subset.

[0031] Preferably, the number of features and the contribution of each subset are determined, and each subset is filtered based on the contribution, background co-response redundancy, and number of features to obtain multiple non-dominated subsets. To construct a set of non-dominated subsets, including: The subset contribution of each subset in the valid candidate set is calculated using the feature information contribution evaluation function. And determine the number of features for each subset in the valid candidate set. ,and ,in, Representing a subset The number of features; Construct filtering rules, and perform pairwise comparisons on all subsets in the valid candidate set based on the constructed filtering rules, and filter out multiple non-dominated subsets based on the comparison results; The filtering rules include: subset with subset As two subsets to be compared, the subsets with subset The subset contribution, background co-response redundancy, and number of features are compared, and the following relationship is satisfied: ; ; ; Furthermore, if at least one inequality holds, then the subset is... Divide into non-dominated subsets; After comparing all subsets in the valid candidate set according to the filtering rules, all non-dominated subsets are constructed into a non-dominated subset set.

[0032] In the above embodiments, the subset contribution of all subsets in the effective candidate set is calculated using a feature information contribution evaluation function, and the number of features in each subset is determined, thus quantifying the three objectives of information content, background co-response redundancy, and compactness. A screening rule containing three inequalities is constructed, and all subsets in the effective candidate set are compared pairwise. A subset is classified as non-dominated if it is not inferior to other subsets in subset contribution, not superior in background co-response redundancy, and not more than other subsets in number of features, and at least one inequality is strictly true. This definition of dominance formalizes the incommensurability between the three objectives. All dominated and non-dominated subsets are constructed as independent sets, allowing subsequent state selection to choose the best from the non-dominated frontier, rather than simply ranking by a single indicator.

[0033] In this embodiment, the screening rule compares the subset contribution, background co-response redundancy, and feature count of two subsets. The subset contribution is calculated using a constructed feature contribution evaluation function. The background co-response redundancy is calculated using the difference between single feature contribution and joint feature contribution. The feature count is the number of features in the subset. When comparing using the screening rule, subset contribution is compared with subset contribution, background co-response redundancy with background co-response redundancy, and feature count with feature count. The subsets with the highest contribution, lowest background co-response redundancy, and lowest feature count are used as the criteria to screen and divide all subsets in the valid candidate set. with subset For example, only when the following conditions are met: ; ; ; Furthermore, at least one inequality must hold, meaning that not all three comparison terms can be equal, before the subset is changed. Divide into non-dominant subsets; otherwise, do not partition the subsets. with subset Instead of partitioning, a new subset is selected for comparison. When a new subset is selected for comparison, if the subset... with subset The comparison partitioning has been completed, then the subset with subset It will not be selected again, but if the subset with subset If a subset is not selected because it does not meet the partitioning criteria, it is possible that a subset will be selected again during the reselection process. with subset .

[0034] It should be noted that, due to the nonlinear growth characteristic of KL divergence evaluation value that depends on subset size, and the cumulative growth of background co-response redundancy according to the number of feature pairs, there are different scale evolution laws between the two and the number of features. Therefore, the traditional weighted summation comprehensive evaluation scheme cannot establish a stable quantitative compromise relationship among the three. This invention uses Pareto non-dominated ranking to replace weighted summation, which fundamentally avoids the allocation of weights to indicators with different dimensions that have no physical meaning. This makes the screening results unaffected by the scale differences of the evaluation indicators and only depends on the objective dominance relationship between subsets.

[0035] Preferably, an optimal feature subset is constructed. The subset with the largest contribution from the non-dominated subset set is selected as the initial feature for the next iteration and updated to the optimal feature subset. The non-dominated subset set is obtained by iterative processing based on the initial feature for the next iteration. ,include: The subset contributions of all subsets in the non-dominated subset set are compared, and the subset with the largest contribution is selected as the optimal feature subset. This optimal feature subset is then used as the initial feature for the next iteration, where the initial feature for the next iteration satisfies the following relationship: ; In the formula, This represents the initial features in the next iteration. The subset with the largest contribution. The corresponding subset contribution; Construct an optimal feature subset, and update the optimal feature subset to the optimal feature subset, wherein the update to the optimal feature subset satisfies the following relationship: ; In the formula, Represents the optimal feature subset; A new set of valid candidates is constructed based on the initial features of the next iteration, and a set of non-dominated subsets is determined from the set of valid candidates.

[0036] In the above embodiments, by comparing the subset contributions of all subsets in the non-dominated subset set, the subset with the largest contribution is selected as the initial feature for the next iteration, ensuring that the discriminative ability continuously improves during the search process. The selected optimal subset is updated to the optimal optimized feature subset, and the globally optimal record is continuously maintained during the iteration process to ensure that even if the search direction temporarily deviates, the final output result can still be traced back to the best combination in the entire search history. Based on the initial features of the next iteration, a new effective candidate set is constructed and the non-dominated subset set is determined, enabling the search process to continuously expand to better regions based on the optimization results of the previous stage.

[0037] Preferably, in each iteration, the optimal feature subset of the current iteration is compared with the optimal optimized feature subset, the optimal optimized feature subset is updated based on the comparison result, and when the iteration meets the termination condition, the updated optimal feature subset is output as the final optimal optimized feature subset, including: The subset contribution of the current iteration's optimal feature subset is compared with the subset contribution of the optimal optimized feature subset. If the subset contribution of the current iteration's optimal feature subset is greater than the subset contribution of the optimal optimized feature subset, then the current iteration's optimal feature subset is updated to the optimal optimized feature subset. If the subset contribution of the current iteration's optimal feature subset is less than the subset contribution of the optimal optimized feature subset, then the optimal feature subset remains unchanged. A set of visited states is constructed based on the initial features from all iterations, where the set of visited states satisfies the following relationship: ; In the formula, Indicates initial features; When constructing a new set of valid candidates, it is determined whether a subset of the valid candidate set exists in the set of visited states, and the subset that exists in the set of visited states is removed from the valid candidate set. When constructing a new set of valid candidates, if all subsets of the valid candidate set exist in the visited state set, it is determined that a new set of valid candidates cannot be obtained at present. The iteration stops when the maximum number of iterations is reached or no new valid candidate set can be obtained. The subset with the highest contribution is then selected from the non-dominated subset set obtained in the last iteration. ; Select the subset with the largest subset contribution from the set of non-dominated subsets obtained in the last iteration. Subset contribution of the optimal feature subset Compare and, Greater than At that time, Update to the optimal feature subset, in Less than At the same time, the optimal feature subset remains unchanged; The updated optimal feature subset is output as the final optimal feature subset.

[0038] In the above embodiments, by comparing the subset contribution of the current iteration's optimal feature subset with the subset contribution of the globally optimal optimized feature subset at each iteration, and updating only when the former is greater than the latter, the monotonic non-decreasing property of the globally optimal solution is guaranteed throughout the search process, effectively avoiding the loss of discovered optimal combinations due to local perturbations. By constructing a visited state set to record the states of all historically explored subsets, and automatically removing visited subsets when constructing a new effective candidate set, the entire search path is remembered, avoiding repeated oscillations or loops between different paths. When all subsets in the effective candidate set exist in the visited state set, it is determined that no new effective candidate set can be obtained. Combined with the maximum number of iterations, this constitutes a double termination condition, avoiding invalid infinite loops and providing a natural convergence endpoint for the search. At termination, the optimal subset obtained in the last iteration is compared with the globally optimal feature subset, and the one with the greater contribution is output. A secondary confirmation mechanism is introduced at termination to ensure that the final output feature subset is the globally optimal solution verified through all iterations.

[0039] In this embodiment, the final optimal feature subset can be directly used for model training. After obtaining the final optimal feature subset, a corresponding low-dimensional feature input is constructed, and the garbage image classification model is trained or updated accordingly. The garbage image classification model can be a convolutional neural network classifier or other image recognition model, used to classify input garbage images. After the model training is completed, the final optimal feature subset is used as a fixed feature selection scheme to extract and filter features from the image to be classified, and then input into the classification model to output the corresponding garbage category recognition result. Through the above termination judgment and result output mechanism, the system not only completes feature subset optimization, but also realizes the construction of a garbage image classification model based on optimized features, thereby improving classification performance and computational efficiency while reducing feature dimensionality.

[0040] Example 2 This application provides a feature subset optimization method for garbage image classification, including: First, the garbage image is input into a pre-defined convolutional neural network for feature extraction, and the feature map response output by the convolutional layer is obtained. Since convolutional feature maps typically contain multiple visual feature channels, and different channels correspond to different local visual mode responses, the multiple channels in the convolutional feature map are defined as the complete set of visual convolutional features to be optimized. ; in, Represents the first feature map in the convolutional feature map. One visual feature channel, This represents the dimension of image features. Based on this, following the method described in Example 1, the entire set of convolutional visual features is pre-weighted according to the spatial activation concentration of each convolutional visual feature channel to obtain a weighted set of visual features. Any subset of features is represented as... For a given feature subset By performing dimensionality filtering on the image feature vectors, a subspace sample set is obtained: ; in , indicating the first The representation of a garbage image in the selected feature dimension. This indicates the number of image samples.

[0041] Based on this, the present invention employs the kernel density method to characterize the local similarity relationship of image features in a subspace, defining the subspace kernel function as follows: ; Based on this kernel function, a category distribution function induced by feature subsets is constructed. ; Meanwhile, the prior distribution of categories under the featureless condition is defined as follows: ; An evaluation function for feature subsets is constructed based on the above distribution. ; in This represents the Kullback–Leibler divergence, used to measure the divergence within a feature subset. The degree of deviation of the category distribution under given conditions from the prior distribution; the larger the value, the more significant the contribution of the feature subset to the classification of waste.

[0042] In the feature subset search process, this invention employs a multi-objective feature subset optimization strategy based on local neighborhood expansion to iteratively optimize the feature combination space. Initialize the feature set. At that time, each feature in the weighted visual feature set is randomly selected with a preset probability, and the first feature is defined as... The state of the next iteration is .

[0043] In each step of the search, a set of candidate subsets is generated by performing feature addition and deletion operations on the current state. Calculate the information contribution evaluation value for all candidate subsets. .

[0044] Because garbage images commonly exhibit complex background regions, repetitive pollution textures, and similar packaging colors, different convolutional visual channels tend to produce synchronized responses to the same garbage background region or local pollution texture. To suppress the repetitive representation of the same garbage visual pattern by multiple visual channels, this invention defines a background co-response redundancy function based on the marginal contribution change of features to the sub-category distribution. For any feature... Its single feature information contribution is ; For any pair of features Their joint information contribution is ; Based on this, the background co-response redundancy of the candidate subset is defined as the degree to which different visual channels produce repeated responses to the same garbage visual region. ; Based on this, a multi-objective evaluation system is constructed, which simultaneously considers information contribution, garbage category discrimination ability, background co-response suppression ability, and feature subset compactness. First, candidate subsets that meet the conditions for improving information contribution are selected to form an effective set; then, based on the Pareto optimality criterion, a selection is made to obtain non-dominated subsets that achieve a balance between information contribution, background co-response redundancy, and size.

[0045] Within the non-dominated subset, select the feature subset with higher information contribution and relatively compact structure as the next state. And continuously maintain the global optimal solution during the search process. .

[0046] The search process terminates when no valid candidate subset exists or the maximum number of iterations is reached, yielding the optimal feature subset. .

[0047] After obtaining the optimal feature subset, the features of the garbage image are filtered based on the feature subset to construct a low-dimensional feature representation, which is then input into the garbage image classification model for training or updating. The classification model classifies the target region in the sub-input garbage image based on the filtered low-redundancy visual features, and reduces the interference of background areas, polluted areas, and repetitive textures on the classification decision, thereby achieving automatic classification of garbage images.

[0048] Example 3 Based on Embodiment 1, this application provides a feature subset optimization method based on spatial activation concentration pre-weighting.

[0049] In garbage images, the garbage target area typically exhibits a locally concentrated visual pattern, while background areas such as conveyor belts, container walls, and the ground usually occupy most of the image area and have relatively uniform texture. In the feature map of a convolutional neural network, the visual feature channels that respond to the garbage target area typically have high activation values ​​concentrated in local locations within the feature map space; while the visual feature channels that respond to the background area typically have activation values ​​that are uniformly distributed or widely dispersed across the entire feature map space. Based on this physical geometric characteristic, this embodiment introduces a spatial activation concentration pre-weighting step after convolutional feature extraction and before feature subset search in Embodiment 1.

[0050] Specifically, after performing convolutional feature extraction using a pre-defined image feature extraction model, the feature map output by the last convolutional module is obtained, with a dimension of 1. ,in and These represent the spatial height and width of the feature map, respectively. This represents the number of channels in the feature map, i.e., the dimension of the entire set of convolutional visual features. For each channel... Extract the channel in Calculate the spatial activation mean of the activation matrix in space. Spatial activation variance : ; ; In the formula, Indicates the first The spatial location of each channel The activation value at that location. The activation concentration of that channel is calculated based on the spatial activation mean and variance. : ; In the formula, For a preset small constant (such as This is used to prevent division-by-zero anomalies in channels with a mean of zero. The physical meaning of activation concentration is: when the activation response of a channel is concentrated in a local garbage target region, high activation values ​​are concentrated in a few spatial locations, and the spatial variance... Larger and spatial mean Smaller, therefore The value is relatively high; when the activation response of the channel evenly covers the background area, the activation values ​​at each spatial location are similar, and the spatial variance is low. Smaller and spatial mean Larger, therefore The value is low.

[0051] For all The activation concentration of each channel is normalized to obtain the spatial concentration weight of each channel. : ; Then, the original feature values ​​of each channel obtained after global average pooling are... After weighted adjustment, the weighted eigenvalues ​​are obtained. : ; In the formula, The preset weight scaling factor, The weighted formula uses In the form of, rather than by directly multiplying The purpose is to ensure that even the background scattering channels with the lowest spatial concentration are not completely zeroed out—because although such channels exhibit background responses in the geometric domain, they may still carry information with marginal discriminative value regarding waste category distribution at the information theory level (e.g., the coexistence relationship between a specific waste category and a specific background environment), and geometric domain features alone are insufficient to completely exclude their discriminative contribution. Weight scaling factor Controlling the overall intervention intensity of the geometric domain pre-weighting: The larger the value, the stronger the pre-screening power of the geometric domain information on the feature space; when When the weighted state is reduced to an unweighted state, it is easier to conduct ablation experiments to verify the technical effect of the preweighted weighting.

[0052] Weighted eigenvalues ​​of all channels ( This constitutes the complete set of weighted visual features. This serves as the input space for subsequent iterative search of feature subsets, replacing the original complete set of convolutional visual features.

[0053] It should be noted that spatial activation concentration captures the activation distribution characteristics of convolutional channels in geometric space, reflecting the spatial structural attribute of "where the feature channels respond in the image"; while the subsequent KL divergence feature information contribution evaluation function captures the class discrimination contribution of feature subsets in probability distribution space, reflecting the information theory attribute of "how much perturbation the class distribution is caused by the feature channels". These two functions operate in the geometric and information domains respectively, belonging to two orthogonal physical dimensions: two channels with the same spatial concentration may have completely different class discrimination capabilities, and vice versa. Therefore, they cannot be derived from or substituted for each other.

[0054] By pre-weighting the concentration in the geometric domain before the KL divergence evaluation in the information domain, the initial feature space entering subsequent iterative evaluations and Pareto screening processes has undergone physical geometric pre-discrimination of garbage targets and background, rather than relying solely on statistical distribution-level category differences for identification. This dual-domain collaborative strategy effectively reduces the search space size for calculating each candidate subset in subsequent KL divergence evaluations and pairwise comparisons in Pareto—low-concentration channels have lower initial weighting values ​​and tend to be excluded from the effective candidate set or eliminated by non-dominated subsets in the early stages of iterative search, thereby reducing the overall computational cost of iterative search and improving the geometric discriminative focusing ability of the final screening results on garbage target regions.

[0055] In the specific experiments of this embodiment, the following experimental environment was configured: On the hardware side, a high-performance computing server was used as the experimental platform, with an Intel Core 12700 processor, 32GB of memory, and a Tesla V100 graphics processing unit with 16GB of video memory to accelerate feature processing and model training. On the software side, a Windows 11 operating system was used, Python 3.10 was used as the programming language, the scikit-learn library was used to implement traditional classification models such as support vector machines, the PyTorch framework was used to implement neural network models, and NumPy and SciPy were combined to complete matrix operations and data processing. Regarding the dataset, the publicly available waste classification dataset TrashNet was selected as the experimental data source. This dataset contains image samples from multiple categories such as paper, plastic, glass, metal, and other waste, characterized by small visual differences between categories and complex backgrounds, effectively validating the performance of feature selection methods in practical image classification tasks. All images were uniformly scaled to a fixed size and normalized before being input into the model.

[0056] In the feature extraction stage, a ResNet18 convolutional neural network is used to learn feature representations for the garbage image, and the channel response output of the last convolutional module is extracted as the original visual feature representation. Specifically, the input image is scaled to 224×224×3 and then input into the ResNet18 network. The residual convolutional module extracts edge texture, color regions, and local structural information from the garbage image layer by layer. In the last convolutional stage, a one-dimensional feature vector is obtained through global average pooling, where each feature corresponds to the global response value of one convolutional channel. Therefore, the resulting vector dimension is used as the complete set of visual convolutional visual features. .

[0057] After obtaining the image feature representation, the feature vector is input into the feature selection module of this invention to filter the high-dimensional feature set and obtain the optimal feature subset. The kernel function adopts a Gaussian kernel form, and its bandwidth parameter is taken as the standard deviation of the pairwise Euclidean distance between samples in the subspace, which is adaptively determined according to the feature distribution of the subspace; the maximum number of iterations in the search parameters is set to... .

[0058] To verify the effectiveness of the feature subsets, a control experiment was constructed during the classification stage. This included five groups of methods: Baseline group: Classified using all features extracted by ResNet18; Traditional methods: Employ mutual information-based feature selection methods to filter features; This embodiment's method group (specifically includes three different implementations): ① without background co-response redundancy constraints and without spatial activation concentration pre-weighting mechanism, ② without spatial activation concentration pre-weighting mechanism, and ③ the complete solution.

[0059] For the classification model, to maintain experimental consistency, a fully connected classifier integrated with the feature extraction structure was adopted. Specifically, the selected features were input into a classification module containing a fully connected layer (output dimension being the number of classes) and a Softmax function, and training was performed using the cross-entropy loss function. During training, the Adam optimizer was used with a learning rate of 0.001 and a batch size of 32. The average classification accuracy was calculated using 5-fold cross-validation. To reduce the impact of random initialization on the results, each experiment was repeated 10 times, and the average result was taken. The experimental results are shown in Table 1, where the number of features represents the average proportion of the final retained feature dimension to the original feature dimension. Table 1. Accuracy comparison between the baseline group, the conventional method group, and the method group of this embodiment.

[0060] Experimental results show that, under the same network structure, compared with the baseline method using all features, the method of this invention can improve classification accuracy while reducing feature dimensionality. Compared with feature selection methods based on mutual information, this invention, by constructing a feature evaluation mechanism based on distribution differences, can more effectively characterize the impact of features on the category distribution structure, thereby reducing the introduction of redundant features and improving feature expressive power. Furthermore, when the background co-response redundancy constraint is removed in the complete scheme of this invention, the classification accuracy decreases from 0.907 to 0.886, indicating that background co-response redundancy can effectively suppress the synchronous response of different visual channels to background regions and repetitive texture regions, thereby reducing the degree of information overlap between features and improving the discriminative power of feature subsets. Visual analysis of the convolutional activation regions reveals that the features selected by the method of this invention are more concentrated in the garbage target region, while the activation response to the background region is significantly reduced. In addition, under different training partitioning conditions, the performance fluctuation of the method of this invention is small, indicating that it has good stability and generalization ability.

[0061] Experimental results further show that after introducing spatial activation concentration pre-weighting, the classification accuracy improved from 0.907 to 0.913, and the final selected feature ratio further decreased from 65% to approximately 61%. The reason for this is that spatial concentration pre-weighting utilizes the geometric difference in the feature map space—where waste targets exhibit locally concentrated high activation, while the background region exhibits large-area scattered activation—to assign a lower initial weight to the low-concentration background scattering channels at the feature selection entry point. These channels received moderate evaluation values ​​in the subsequent KL divergence evaluation—they showed a statistical correlation with the waste category distribution in the probability distribution domain, but this correlation essentially stemmed from a spurious correlation between a specific waste category and a specific background environment, rather than a response to the visual pattern of the waste target itself. After spatial pre-weighting identifies and weakens these channels through geometric domain information, the subsequent Pareto selection tends to retain visual feature channels with high discriminative focusing ability for the real waste target area, thereby improving classification accuracy while reducing feature dimensionality. Furthermore, since the initial weights of low-concentration channels are reduced overall in the weighted visual feature set, the iterative search converges to a stable non-dominated front within approximately 5 to 6 rounds, saving about 15% of computational overhead compared to not introducing a spatial pre-weighting scheme. This application also provides a feature subset optimization system for garbage image classification, including a processor and a memory. Memory, used to store computer programs; When a processor executes a program stored in memory, it implements any of the steps described in the feature subset optimization method for garbage image classification.

[0062] The aforementioned feature subset optimization system for garbage image classification can implement various embodiments of the aforementioned feature subset optimization method for garbage image classification and achieve the same beneficial effects, which will not be elaborated here.

[0063] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A feature subset optimization method for garbage image classification, characterized in that, include: Image datasets are constructed by acquiring image data of garbage images. Feature extraction and spatial activation feature weighting are performed on the image datasets to obtain a complete set of convolutional visual features that characterize the color texture, edge structure, and local morphological information of garbage areas. The complete set of convolutional visual features is pre-weighted based on the spatial activation concentration of each convolutional visual feature channel to obtain a weighted complete set of visual features. Set a maximum number of iterations, randomly select a sub-feature from the weighted visual feature set as the initial feature, and perform a single-step transition operation on the initial feature to construct a candidate set. Then, determine the effective candidate set from the candidate set using the constructed feature information contribution evaluation function, including: For any subset in the candidate set, the similarity of the subset in the sample space is calculated using a Gaussian kernel function, where the Gaussian kernel function satisfies the following relationship: ； In the formula, Both indicate that the sample is in the feature subset The projection vector on, Denotes the Euclidean norm. This represents the kernel bandwidth parameter of the corresponding subspace; The label distribution of arbitrary subsets is constructed based on the Gaussian kernel function, where the label distribution satisfies the following relationship: ； In the formula, Indicating in the feature subset Conditional Labels The probability estimate, This is an indicator function; it takes the value 1 when the indicator function condition is true, and 0 otherwise. Indicates the first The feature labels corresponding to each sample Indicates sample and In feature subset projection and Spatial similarity; Construct the label prior distribution under featureless conditions, where the label prior distribution under featureless conditions satisfies the following relationship: ； In the formula, This represents the label distribution without using any feature information. Indicates the first The feature labels corresponding to each sample; A feature information contribution evaluation function is constructed based on the label distribution of arbitrary subsets and the prior label distribution under featureless conditions. The feature information contribution evaluation function satisfies the following relationship: ； ； In the formula, Indicates the Kullback–Leibler divergence. Representing a feature subset The intensity of the distribution disturbance introduced relative to the state without information; The first feature information contribution evaluation function value of the initial features is calculated based on the feature information contribution evaluation function. The evaluation function value of the contribution of the second feature information of all subsets in the candidate set. ; Filter all contribution evaluation function values ​​of the second feature information The value of the contribution evaluation function of the first feature information is greater than A subset of is used to construct a valid candidate set, wherein the construction of the valid candidate set satisfies the following relation: ； In the formula, Denotes the set of valid candidates. Denotes a subset of the candidate set. Denotes the candidate set, Indicates the set that has been visited; The single feature information contribution and joint information contribution of each subset in the effective candidate set are calculated based on the feature information contribution evaluation function, and the background co-response redundancy of each subset is calculated based on the degree of deviation between the single feature information contribution and the joint information contribution. The number of features and contribution of each subset are determined. Based on the contribution, background co-response redundancy and the number of features, each subset is filtered to obtain multiple non-dominated subsets to construct a set of non-dominated subsets. Construct an optimal feature subset, select the optimal feature subset from the non-dominated subset set as the initial feature of the next iteration process and update it to the optimal feature subset, and iterate based on the initial feature of the next iteration process to obtain the non-dominated subset set; In each iteration, the best feature subset of the current iteration is compared with the optimal optimized feature subset. The optimal optimized feature subset is updated based on the comparison result. When the iteration meets the termination condition, the updated best feature subset is output as the final optimal optimized feature subset to classify garbage images.

2. The feature subset optimization method for garbage image classification according to claim 1, characterized in that, The process involves acquiring image data from garbage images to construct an image dataset, and then performing feature extraction and spatial activation feature weighting on the image dataset to obtain a complete set of convolutional visual features representing the color texture, edge structure, and local morphological information of the garbage region. The complete set of convolutional visual features is then pre-weighted based on the spatial activation concentration of each convolutional visual feature channel to obtain a weighted complete set of visual features, including: By acquiring garbage images and performing a unified modeling on them, a garbage image dataset is obtained, which satisfies the following relationship: ； In the formula, Represents an image dataset. Indicates the first A sample of garbage images, This indicates the corresponding waste category label. Indicates the total number of samples; The image feature extraction model extracts the mapped image from the image dataset. The feature representation of the 1D feature vector yields the corresponding complete set of convolutional visual features, wherein the complete set of convolutional visual features satisfies the following relationship: ； In the formula, Represents the complete set of visual features from convolution. Representing image features, Indicates the dimension of image features; Obtain the output of the last convolutional module of the convolutional neural network 3D feature map, where These represent the spatial height and width of the feature map, respectively. Number of channels; For each channel Calculate the feature map of this channel in The activation concentration in space, which satisfies the following relationship: ； ； ； In the formula, Indicates the first The spatial location of each channel Activation value at that location, This represents the spatial activation mean of the channel. This represents the spatial activation variance of the channel. A preset small constant to prevent division by zero; Channel weights are generated based on the activation concentration of each channel, and the feature values ​​of each channel after global average pooling are weighted. The channel weights and weighted feature values ​​satisfy the following relationships: ； ； In the formula, Represents the normalized i-th Each channel weight, Indicates the weighted number of... Each channel feature value, Represents the first digit after the original global average pooling. Each channel feature value, This represents the preset weight scaling factor. ; All weighted channel feature values ​​constitute the complete set of weighted visual features.

3. The feature subset optimization method for garbage image classification according to claim 1, characterized in that, The step of randomly selecting a sub-feature from the entire set of weighted visual features as an initial feature, and constructing a candidate set by performing a single-step transition operation on the initial feature, includes: Randomly select a sub-feature from the entire set of weighted visual features as the initial feature, and define the initial feature to satisfy the following relationship: ； In the formula, Indicates initial features, This represents a feature index randomly selected from the entire set of weighted visual features. Indicates the maximum number of iterations; The single-step transition operation includes an addition operation and a deletion operation. By performing the addition and deletion operations in the single-step transition operation on the initial feature, a neighborhood candidate set of the current state is generated. The add operation includes: Each sub-feature in the weighted visual feature set is used as an initial feature. For each initial feature, if the corresponding initial feature does not exist in the current feature subset state... When the initial feature is added to the current feature subset state, a neighborhood set is generated. The neighborhood set satisfies the following relationship: ； In the formula, Indicates the first The next iteration corresponds to adding to the neighborhood set. Represents the first element in the weighted visual feature set. Individual characteristics; Deletion operations include: For each initial feature, the corresponding initial feature is removed from the current feature subset state, generating a deletion neighborhood set. The deletion neighborhood set satisfies the following relation: ； In the formula, Indicates the first The deleted neighborhood set corresponding to the next iteration Represents the first element in the weighted visual feature set. Individual characteristics; A candidate set is constructed based on adding to and deleting from the neighborhood set. The candidate set satisfies the following relationship: ； In the formula, Indicates the first The candidate set generated in the next iteration.

4. The feature subset optimization method for garbage image classification according to claim 1, characterized in that, The single feature information contribution and joint information contribution of each subset in the effective candidate set are calculated based on the feature information contribution evaluation function, including: The single feature information contribution of each feature in all subsets of the effective candidate set is calculated based on the feature information contribution evaluation function. ,and, ; The joint information contribution of any two features within any subset containing two or more features in all subsets of the valid candidate set is calculated based on the feature information contribution evaluation function. ,and, .

5. The feature subset optimization method for garbage image classification according to claim 4, characterized in that, The background co-response redundancy of each subset is calculated based on the deviation between the single feature information contribution and the joint information contribution, including: The background co-response redundancy of all subsets of the effective candidate set is calculated based on the difference between the single feature information contribution and the joint information contribution, wherein the calculation of the background co-response redundancy satisfies the following relationship: ； In the formula, Represents any feature subset Background co-response redundancy, , Both represent contributions from single feature information. This represents the joint feature information contribution of the two features.

6. The feature subset optimization method for garbage image classification according to claim 5, characterized in that, The number of features and contribution of each subset are determined. Based on the subset contribution, the background co-response redundancy, and the number of features, each subset is filtered to obtain multiple non-dominated subsets to construct a non-dominated subset set, including: The subset contribution of all subsets in the effective candidate set is calculated using the feature information contribution evaluation function. And determine the number of features in all subsets of the valid candidate set. ,and ,in, Representing a subset The number of features; Build filtering rules, and perform pairwise comparisons on all subsets in the valid candidate set based on the built filtering rules, and filter out multiple non-dominated subsets based on the comparison results; The filtering rules include: subset with subset As two subsets to be compared, the subsets with subset The subset contribution, background co-response redundancy, and number of features are compared, and the following relationship is satisfied: ； ； ； Furthermore, if at least one inequality holds, then the subset is... As a non-dominated subset; After comparing all subsets in the valid candidate set according to the filtering rules, all non-dominated subsets are constructed into a non-dominated subset set.

7. The feature subset optimization method for garbage image classification according to claim 6, characterized in that, Construct an optimal feature subset, select the optimal feature subset from the non-dominated subset set as the initial feature for the next iteration and update it to the optimal feature subset, and iterate based on the initial feature for the next iteration to obtain the non-dominated subset set, including: The subset contributions of all subsets in the non-dominated subset set are compared, and the subset with the largest contribution is selected as the optimal feature subset. This optimal feature subset is then used as the initial feature for the next iteration, where the initial feature for the next iteration satisfies the following relationship: ； In the formula, This represents the initial features in the next iteration. The subset with the largest contribution. The corresponding subset contribution; Construct an optimal feature subset, and update the optimal feature subset to the optimal feature subset, wherein the update to the optimal feature subset satisfies the following relationship: ； In the formula, Represents the optimal feature subset; A new set of valid candidates is constructed based on the initial features of the next iteration, and a set of non-dominated subsets is determined from the set of valid candidates.

8. The feature subset optimization method for garbage image classification according to claim 7, characterized in that, In each iteration, the optimal feature subset of the current iteration is compared with the optimal optimized feature subset. Based on the comparison result, the optimal optimized feature subset is updated. When the iteration meets the termination condition, the updated optimal feature subset is output as the final optimal optimized feature subset, including: The subset contribution of the current iteration's optimal feature subset is compared with the subset contribution of the optimal optimized feature subset. If the subset contribution of the current iteration's optimal feature subset is greater than the subset contribution of the optimal optimized feature subset, then the current iteration's optimal feature subset is updated to the optimal optimized feature subset. If the subset contribution of the current iteration's optimal feature subset is less than the subset contribution of the optimal optimized feature subset, then the optimal feature subset remains unchanged. A set of visited states is constructed based on the initial features from all iterations, where the set of visited states satisfies the following relationship: ； In the formula, Indicates initial features; When constructing a new set of valid candidates, it is determined whether a subset of the set of valid candidates exists in the set of visited states, and the subset that exists in the set of visited states is removed from the set of valid candidates. When constructing a new set of valid candidates, if all subsets of the valid candidate set exist in the visited state set, it is determined that a new set of valid candidates cannot be obtained at present. The iteration stops when the maximum number of iterations is reached or no new valid candidate set can be obtained. The subset with the highest contribution is then selected from the non-dominated subset set obtained in the last iteration. ; Select the subset with the largest subset contribution from the set of non-dominated subsets obtained in the last iteration. Subset contribution of the optimal feature subset Compare and, Greater than At that time, Update to the optimal feature subset, in Less than At the same time, the optimal feature subset remains unchanged; The updated optimal feature subset is output as the final optimal feature subset.

9. A feature subset optimization system for garbage image classification, characterized in that, Including processor and memory; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method as described in any one of claims 1-8.

Images