Clothes recognition method and system for intelligent clothes care machine

Abstract

This invention discloses a clothing recognition method and system for an intelligent clothing care machine, belonging to the field of image processing and intelligent control technology. The method involves acquiring input image data of the clothing to be recognized and extracting initial recognition feature vectors using an image recognition model. Based on these feature vectors, the image is segmented to extract local texture features and edge contour features for each candidate clothing region. These features are then input into a material discrimination sub-model, which outputs corresponding material labels. Image features and material labels are fused to generate a fused feature vector, which is then input into a semantic enhancement model for iterative optimization to obtain semantically enhanced recognition features. These features are then input into a joint clothing type and care need recognition model, which outputs clothing classification results and recommended care modes, driving the intelligent clothing care machine to perform corresponding care tasks. This invention achieves high-precision recognition of multi-dimensional clothing features and personalized care recommendations, with high recognition accuracy and a high degree of intelligence.

Description

Technical Field

[0001] This invention relates to the fields of image processing and intelligent control technology, specifically to a clothing recognition method and system for intelligent clothing care machines. Background Technology

[0002] Against the backdrop of the rapid development of smart home appliances, smart garment care machines, as composite home appliances integrating washing, drying, ironing, and care, are increasingly widely used in homes and commercial spaces. Due to their complex functions, these devices need to accurately identify the type, material, color, and care requirements of the garments placed inside in order to automatically select the appropriate care program. However, most existing garment recognition technologies rely on single-dimensional features (such as color or texture) for preliminary classification, resulting in low accuracy. This is especially problematic in real-world scenarios where various garments are mixed together, lighting conditions change, or garments partially obscure the clothing, leading to recognition errors, inappropriate selection of care modes, and consequently, problems such as garment deformation, fading, or damage.

[0003] More seriously, for certain special garments (such as those with functional coatings, heat-sensitive materials, or three-layer composite structures), existing identification systems cannot distinguish their internal structures, which may look similar but have completely different care requirements. They are easily misidentified as ordinary garments, which can damage the garments during high-temperature steam or high-speed spin drying, seriously affecting the user experience and the practical value of the equipment. Summary of the Invention

[0004] The purpose of this invention is to provide a clothing recognition method and system for intelligent clothing care machines to address the shortcomings in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a clothing recognition method for an intelligent clothing care machine, comprising: S100: Obtain the image input data of the clothing to be identified and input it into the preset image recognition model M0 to obtain the initial recognition feature vector F0; S110. Based on the initial recognition feature vector F0, the input image is segmented to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions, and the local texture features Lt and edge contour features Et of each region are extracted respectively. S120. Input the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and output the material discrimination labels T1~Tn corresponding to each region. S130. The initial identification feature vector F0 is fused with all material labels T1~Tn to generate a fused feature vector Ff; S140. The fused feature vector Ff is iteratively optimized using the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe. S150. Input the semantically enhanced recognition feature vector Fe into the clothing type and nursing needs joint recognition model M3, and output the classification result of the target clothing and the recommended nursing mode. S160. Based on the classification results and recommended care modes, the corresponding care tasks are performed using the intelligent clothing care machine.

[0006] Preferably, S100 includes: S101. Obtain a multi-angle image sequence of the clothing to be identified under natural light or artificial light source, including front, back and side views; S102. Perform image preprocessing on the image sequence; S103. Input the preprocessed image sequence into the image recognition model M0, extract the corresponding local texture features, global contour features and color histogram features to form the initial recognition feature vector F0.

[0007] Preferably, S110 includes: S111. Based on the color distribution and texture gradient information contained in the initial recognition feature vector F0, the superpixel clustering algorithm is used to perform preliminary segmentation of the input image to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions. S112. The candidate region is optimized by using a graph cut-based image segmentation method. S113. For each optimized candidate clothing region, extract its gray-level co-occurrence matrix to construct local texture features Lt, and use the Canny algorithm to extract edge contour features Et. S114. Jointly encode the local texture features Lt and edge contour features Et of each candidate region.

[0008] Preferably, wherein S120 includes: S121. Perform dimensional unification and normalization on the local texture features and edge contour features of each candidate region to construct a standardized input feature vector; S122. Input the standardized input feature vector into the pre-trained material discrimination sub-model M1. The material discrimination sub-model M1 is constructed based on the support vector machine algorithm, and the radial basis kernel function is used to optimize the parameters through cross-validation. S123. Perform maximum value judgment on the probability distribution results of each category output by the material discrimination sub-model M1 to obtain the corresponding material discrimination label; S124. Bind the material discrimination label to the candidate area index to generate a one-to-one correspondence between candidate clothing areas and material categories.

[0009] Preferably, wherein S130 includes: S131. Perform one-hot encoding on the material tag corresponding to each candidate clothing area to generate a fixed-dimensional material vector group. S132. Concatenate and splice the initial identification feature vector F0 and the material vector group in the order of regions to form a candidate region-level fusion feature matrix. S133. Use principal component analysis algorithm to perform feature dimensionality reduction on the fused feature matrix and retain the principal components with a cumulative contribution rate greater than 95%. S134. Perform global average pooling on the dimensionality-reduced fusion feature matrix to generate the fusion feature vector Ff.

[0010] Preferably, S140 includes: S141. Input the fused feature vector Ff into the multi-layer attention-aware network of the image semantic enhancement model M2 to extract the initial semantic context features; S142. Utilize a bidirectional feature residual connection structure to perform deep iterative optimization of the initial semantic context features; S143. Adjust the weights of each feature dimension to increase the proportion of key semantic features in the output vector; S144. The optimized semantic context features are normalized and activated to output the semantic enhancement recognition feature vector Fe.

[0011] Preferably, wherein S150 includes: S151. Input the semantically enhanced recognition feature vector into the dual-channel classification structure of the clothing type and nursing needs joint recognition model, and extract clothing category features and nursing feature sub-vectors respectively. S152. The clothing category feature sub-vectors are processed by a fully connected layer and a Softmax function to output the classification result of the target clothing. S153. Input the nursing feature sub-vectors into the multi-label output layer, and combine them with the multilayer perceptron to calculate the fit probability of each nursing option. S154. Based on the classification results and the matching probability, select the nursing option with the highest probability as the recommended nursing mode, and combine it with the clothing category results for output.

[0012] The present invention also provides a garment recognition system for a smart garment care machine, comprising: The image acquisition module acquires the image input data of the clothing to be identified and inputs it into the preset image recognition model M0 to obtain the initial recognition feature vector F0. The feature extraction module performs region segmentation on the input image based on the initial recognition feature vector F0 to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions, and extracts the local texture features Lt and edge contour features Et of each region respectively. The material recognition module inputs the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and outputs the material discrimination labels T1~Tn corresponding to each region. The feature fusion module fuses the initial recognition feature vector F0 with all material labels T1~Tn to generate a fused feature vector Ff; The semantic enhancement processing module iteratively optimizes the fused feature vector Ff through the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe. The joint identification module inputs the semantically enhanced identification feature vector Fe into the clothing type and nursing needs joint identification model M3, and outputs the classification result of the target clothing and the recommended nursing mode. The nursing task execution module executes the corresponding nursing tasks through the intelligent clothing care machine based on the classification results and recommended nursing modes.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention constructs a multi-level recognition system consisting of an image recognition model, a material discrimination sub-model, a semantic enhancement model, and a joint recognition model. This system enables high-precision recognition and classification of mixed clothing in real-world scenarios involving complex stacking, occlusion, and similar materials. Compared to traditional recognition methods that rely on single image features or shallow classification models, this invention integrates local texture, edge contours, and material label information. Through semantic enhancement and a dual-channel recognition strategy, it effectively improves the accuracy and generalization ability of recognition, significantly reduces the recognition error rate, and enhances the model's adaptability to diverse clothing characteristics.

[0014] 2. Based on the recognition results, this invention further links and matches clothing type with recommended care modes, and deeply integrates this with the control logic of intelligent clothing care equipment, achieving a fully automated closed-loop process from image input to care task execution. It automatically adjusts care program parameters according to the material characteristics and care requirements of different garments, reducing user intervention and avoiding damage to clothing due to improper operation. This invention significantly improves the automation level and personalized processing capabilities of intelligent clothing care equipment, possessing strong practical value and promising industrial application prospects. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, the clothing recognition method for a smart clothing care machine includes: S100: Obtain the image input data of the clothing to be identified and input it into the preset image recognition model M0 to obtain the initial recognition feature vector F0.

[0020] In this embodiment of the invention, acquiring a multi-angle image sequence of the clothing to be identified under natural or artificial light specifically involves controlling the shooting device to rotate and capture images around the clothing at fixed intervals. The shooting angles include a front view, a back view, and two side views, acquiring at least four still images to form an image sequence. During the shooting process, the ambient lighting should remain stable. When using artificial light, the main light source should be set to a brightness of no less than 1000 lumens, and the color temperature should be controlled between 5000 and 6000 Kelvin to ensure image color reproduction and clarity.

[0021] The steps for image preprocessing of the image sequence include: First, performing brightness normalization processing on each frame of the image, normalizing the image pixel values ​​to a grayscale range of 0 to 255 to eliminate brightness shifts caused by differences in illumination; second, using a median filtering algorithm for noise suppression, with the median filtering window size set to 3x3 to reduce background interference; and finally, enhancing the clothing edge contours using the Canny edge detection algorithm, with the high and low thresholds of the Canny algorithm set to 120 and 60 respectively to ensure edge continuity and feature integrity.

[0022] The preprocessed image sequence is input into the image recognition model M0 to extract local texture features, global contour features, and color histogram features, forming the initial recognition feature vector F0. The image recognition model M0 employs a deep image classification model based on a convolutional neural network, constructed as follows: the input layer accepts image data of size 224 x 224 pixels, followed by 5 convolutional layers and 3 fully connected layers. All convolutional kernels have a size of 3 x 3 and a stride of 1. The activation function is a modified linear unit, and the pooling layers use max pooling.

[0023] Local texture features are extracted from the feature map output by the third convolutional layer of the model. Global contour features are obtained from the portion of the output vector of the penultimate fully connected layer that retains spatial location information. Color histogram features are obtained by dividing the RGB channels of the input image into eight levels and then statistically analyzing them. These three types of features are combined in a weighted ratio of 3:2:1 to generate an initial recognition feature vector F0, which is used as the basic input for subsequent recognition processing.

[0024] S110. Based on the initial recognition feature vector F0, the input image is segmented to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions. The local texture features Lt and edge contour features Et of each region are extracted respectively.

[0025] In the process of region segmentation based on the initial recognition feature vector F0, preferably, a superpixel clustering algorithm is used to perform preliminary segmentation of the input image. Specifically, this includes: based on the color distribution histogram and texture gradient direction features contained in F0, a Simple Linear Iterative Clustering (SLIC) algorithm is used to cluster image pixels in a five-dimensional space, which consists of the red, green, and blue channel values ​​of the image and two-dimensional coordinate positions. The initial number of superpixels is set to 300, the number of iterations is set to 10, and the superpixel size compactness factor is set to 10. The final output contains multiple relatively uniform candidate clothing regions, denoted as R1, R2, ..., Rn, where n is the total number of candidate clothing regions.

[0026] To improve the consistency between candidate clothing regions and real clothing boundaries, a graph-cut-based image segmentation method is further employed for boundary optimization based on the preliminary segmentation results. The specific process is as follows: each candidate region is constructed as an undirected graph model, where each node represents an image pixel, and edge weights are set based on pixel grayscale similarity and spatial adjacency. The minimum-cut maximum-flow algorithm is used to perform a minimum-cut operation on the graph model, redefining the boundaries. The energy function in the graph cut consists of a region term and a boundary term. The region term is constructed based on a color histogram probability model, and the boundary term uses a Gaussian function to evaluate pixel gradient differences, with a smoothing factor set to 0.8.

[0027] After boundary optimization, feature extraction is performed on each candidate clothing region Ri. First, local texture features Lt within the region are extracted using the gray-level co-occurrence matrix (GLCM), including four statistical indicators: energy, contrast, entropy, and correlation. The GLCM is calculated in 0-degree direction with a step size of 1 and 8 gray levels. Second, the Canny edge detection algorithm is applied to extract the edge contour features Et of the region, with high and low thresholds set to 120 and 60, respectively. Dual threshold concatenation and non-maximum suppression methods are used to ensure the integrity and continuity of edge extraction.

[0028] Finally, the local texture features Lt and edge contour features Et of each candidate region are jointly encoded to construct a fused feature vector, which serves as the input to the subsequent material discrimination model. The joint encoding method is feature vector-level concatenation. First, Lt and Et are normalized so that the feature values ​​of each dimension are distributed between 0 and 1. Then, they are concatenated in a fixed order to ensure the consistency and comparability of the feature space.

[0029] S120. Input the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and output the material discrimination labels T1~Tn corresponding to each region.

[0030] In this embodiment of the invention, the local texture features and edge contour features of each candidate clothing region are first subjected to dimensionality unification and normalization processing. The local texture features consist of a four-dimensional vector derived from the gray-level co-occurrence matrix, comprising energy, contrast, entropy, and correlation. The edge contour features are two-dimensional features extracted based on the Canny algorithm, consisting of edge length density and edge direction histogram. The above six-dimensional feature vectors are concatenated in a unified order, and the maximum-minimum normalization method is used to map each dimension value to the interval between 0 and 1 to construct a standardized input feature vector.

[0031] The standardized input feature vectors are fed into a pre-trained material discrimination sub-model M1, which is constructed based on a support vector machine (SVM) algorithm. The SVM uses a radial basis function (RBF) kernel with a kernel parameter γ set to 0.1 and a penalty factor C set to 100. During model training, a 10-fold cross-validation strategy is used on a standard dataset containing five material categories: cotton, wool, silk, polyester, and leather, with a total of 10,000 samples. The training accuracy remains stable above 93%. A grid search method is used to fine-tune the parameters during training, and the parameter combination with the smallest error is finally selected for model fixing.

[0032] In the actual identification process, the material discrimination sub-model M1 outputs a five-dimensional probability distribution for each input feature vector, corresponding to the matching probability of the five material categories. This invention employs a maximum value decision strategy for this probability distribution result, selecting the category with the highest probability value as the final material discrimination label for the corresponding region. For example, when the model output is "cotton: 0.1, wool: 0.3, silk: 0.05, polyester: 0.15, leather: 0.4", the region is labeled as "leather".

[0033] The material labels obtained from the above discrimination are bound to the corresponding candidate region numbers to generate a one-to-one data structure. Each set of data includes three items: region index number, standardized input feature vector, and material discrimination label, which are used for subsequent feature fusion generation and feature matching calculations in the intelligent recognition path.

[0034] S130. The initial recognition feature vector F0 is fused with all material labels T1~Tn to generate a fused feature vector Ff, which is then passed as input to the image semantic enhancement model M2.

[0035] One-hot encoding is performed on the material tag corresponding to each candidate clothing area. Specifically, assuming there are five predefined material categories: cotton, wool, silk, polyester, and leather, each material tag can be encoded as a five-dimensional vector, where the position corresponding to the category is set to 1, and the remaining positions are set to 0. For example, when a region is identified as "wool," its one-hot encoding result is [0,1,0,0,0]. The one-hot encoding results of all candidate regions are arranged sequentially according to the region number, forming a fixed-dimensional material vector group.

[0036] The initial recognition feature vector F0 is concatenated with the aforementioned material vector group. F0 is the comprehensive feature vector extracted by the image recognition model, and the material vector group is a set of encoded material labels for multiple regions. The concatenation method is as follows: according to the numbering order of the candidate regions, the material vector corresponding to each region is inserted after the feature at the corresponding position in F0, ensuring a one-to-one correspondence of the fused features in spatial structure. After concatenation, a region-level fused feature matrix is ​​formed, where each row of the matrix corresponds to the comprehensive descriptive information of a candidate region, including its image visual features and material semantic features.

[0037] To reduce feature dimensionality, remove redundant information, and improve subsequent processing efficiency, principal component analysis (PCA) is used to reduce the dimensionality of the fused feature matrix. The dimensionality reduction process includes: first, calculating the covariance matrix of the fused feature matrix; then, performing eigenvalue decomposition on the covariance matrix to extract eigenvalues ​​and their corresponding eigenvectors; and finally, sorting the eigenvalues ​​by size and selecting the top principal components whose cumulative contribution rate is greater than 95%. In this embodiment, the input feature matrix has approximately 200 dimensions, which is reduced to less than 40 dimensions after dimensionality reduction, while effectively preserving the joint information structure of the image and material.

[0038] Global average pooling is performed on the dimensionality-reduced fusion feature matrix to further integrate the feature representations between regions and generate the final fusion feature vector Ff. Global average pooling is calculated by taking the arithmetic mean of the values ​​of all regions in the matrix across each feature dimension, outputting a fixed-length fusion vector. Ff, as a highly compressed feature representation possessing both image and semantic information, is passed as input to the subsequent image semantic enhancement model for high-level recognition and classification processing.

[0039] S140. The fused feature vector Ff is iteratively optimized using the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe.

[0040] The input Ff is fed into the image semantic enhancement model M2 for feature optimization. The image semantic enhancement model M2 is constructed using a multi-layer attention-aware network, containing three sequentially stacked attention enhancement sub-layers. Each sub-layer includes a multi-head self-attention module and a feedforward network. The multi-head self-attention module is used to extract the semantic relationships between regions in the input feature vector. The number of heads is set to 8, and the dimensions of keys, values, and queries are all 64. Attention weights are calculated through dot products, and the output is the initial semantic context features.

[0041] Based on the initial semantic context features, a bidirectional feature residual connection structure is used for deep iterative optimization. This residual structure comprises two parallel paths: a forward path that preserves the input features, and a reverse path that performs feature backprojection and reconstruction. The outputs of the two paths are then weighted and fused after dimensional alignment, with the fusion weights automatically adjusted using learnable parameters. Within this structure, a modified linear unit is used as the activation function, and batch normalization is employed for standardization. The optimization process iterates for two rounds, with the updated features from each round serving as the input for the next round.

[0042] After residual optimization, to enhance the expression of key semantic information, a channel attention mechanism is further introduced to adjust the weights of each feature dimension. This channel attention mechanism is based on the statistical importance of features channel by channel and calculates the response weight of each channel through a two-layer perceptual network. The channel compression ratio of the first layer is set to 16, and the sigmoid function is used as the activation function. This achieves dynamic weighting of features along the channel dimension, increasing the proportion of discriminative semantic features in clothing recognition within the overall output vector.

[0043] The optimized semantic context features are then normalized and activated to obtain the final semantically enhanced recognition feature vector Fe. The normalization operation employs a layer-level normalization method, subtracting the mean from each feature dimension and dividing by the standard deviation to improve the stability of the features under different input conditions. The activation process uses a modified linear unit function to increase non-linear expressive power.

[0044] S150. Input the semantically enhanced recognition feature vector Fe into the clothing type and nursing needs joint recognition model M3, and output the classification result of the target clothing and the recommended nursing mode.

[0045] The semantically enhanced recognition feature vectors are input into the joint clothing type and nursing need recognition model M3. This model employs a dual-channel classification structure, consisting of two parallel feature extraction branches, used to extract feature sub-vectors for clothing categories and nursing need features, respectively. Both branches consist of two fully connected layers with modified linear units (MRUs) as the inter-layer activation function. The output dimensions are set to the number of clothing categories and the number of nursing need options, respectively, to achieve decoupled feature representation.

[0046] The clothing category feature vector is fed into a fully connected layer and then processed by a Softmax function. The Softmax function normalizes the scores of each category using an exponential function, outputting a probability distribution vector with a sum of 1. Each dimension represents the probability value of the input sample belonging to a certain clothing category, such as: tops, trousers, skirts, coats, accessories, etc. Finally, the category corresponding to the highest probability is taken as the classification result of the target clothing, and the output result is used for downstream care strategy matching.

[0047] The nursing feature subvectors are input into a three-layer multilayer perceptron for multi-label prediction. The number of nodes in each layer is 128, 64, and 16 respectively. The hidden layers use modified linear units as activation functions, and the output layer uses a sigmoid function, ensuring that each nursing option outputs an independent adaptation probability value. Nursing options include, but are not limited to, drying, steam ironing, low-temperature air drying, deodorization, and sterilization. The model automatically calculates the adaptability of each nursing method based on the input features.

[0048] Finally, by combining the clothing classification results with the compatibility probability of the care options, a recommended care mode output strategy is implemented. In this embodiment, the preferred strategy is to select the care method with the highest compatibility probability as the recommended care mode within the range of care options allowed for the current clothing category. For example, for clothing classified as "wool coat," only the option with the highest probability among "low-temperature air drying," "steam ironing," and "sterilization care" is selected as the final recommended output. The final output includes clothing category labels and care mode suggestions, providing an accurate basis for subsequent automatic program matching of the intelligent clothing care device.

[0049] S160. Based on the classification results and recommended care modes, the corresponding care tasks are performed using the intelligent clothing care machine.

[0050] After obtaining the clothing classification results and recommended care modes, the intelligent clothing care machine automatically matches and executes the corresponding care tasks based on the recognition results. Specifically, the clothing category labels and care suggestions are first input into the task scheduling unit, and the specific execution procedure is determined according to a preset task control mapping table. The mapping table is a pre-built mapping relationship database that records the parameter settings corresponding to each clothing category under a specific care mode, including temperature, humidity, wind speed, care duration, and care action sequence.

[0051] For example, if the identification result is "silk shirt" and the recommended care mode is "low-temperature steam ironing," the system will automatically retrieve the corresponding control parameters: the care temperature is controlled between 40 and 50 degrees Celsius, the steam jet pressure is less than 0.2 MPa, the care duration is set to 5 to 8 minutes, and medium-speed air circulation is maintained during the process to avoid localized moisture. Through the coordinated control of the heater, steam generator, fan, and rotating mechanism inside the care chamber, precise execution of the entire process is achieved.

[0052] During the operation, the garment care machine will also monitor the condition of the clothing and environmental parameters in real time. Using temperature and humidity sensors and clothing tension detection devices, it will determine if any abnormalities are present. If the temperature exceeds a set threshold, the system will immediately terminate the current task and issue an alarm to prompt the user to check the equipment and clothing condition, thus ensuring safety and the quality of garment care.

[0053] Example 2, please refer to Figure 2 As shown, the garment recognition system for the intelligent garment care machine described in this embodiment includes: The image acquisition module acquires the image input data of the clothing to be identified and inputs it into the preset image recognition model M0 to obtain the initial recognition feature vector F0. The feature extraction module performs region segmentation on the input image based on the initial recognition feature vector F0 to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions, and extracts the local texture features Lt and edge contour features Et of each region respectively. The material recognition module inputs the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and outputs the material discrimination labels T1~Tn corresponding to each region. The feature fusion module fuses the initial recognition feature vector F0 with all material labels T1~Tn to generate a fused feature vector Ff; The semantic enhancement processing module iteratively optimizes the fused feature vector Ff through the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe. The joint identification module inputs the semantically enhanced identification feature vector Fe into the clothing type and nursing needs joint identification model M3, and outputs the classification result of the target clothing and the recommended nursing mode. The nursing task execution module executes the corresponding nursing tasks through the intelligent clothing care machine based on the classification results and recommended nursing modes.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A garment recognition method for a smart garment care machine, characterized in that: include: S100: Obtain the image input data of the clothing to be identified and input it into the preset image recognition model M0 to obtain the initial recognition feature vector F0; S110. Based on the initial recognition feature vector F0, the input image is segmented to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions, and the local texture features Lt and edge contour features Et of each region are extracted respectively. S120. Input the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and output the material discrimination labels T1~Tn corresponding to each region. S130. The initial identification feature vector F0 is fused with all material labels T1~Tn to generate a fused feature vector Ff; S140. The fused feature vector Ff is iteratively optimized using the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe. S150. Input the semantically enhanced recognition feature vector Fe into the clothing type and nursing needs joint recognition model M3, and output the classification result of the target clothing and the recommended nursing mode. S160. Based on the classification results and recommended care modes, the corresponding care tasks are performed using the intelligent clothing care machine.

2. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S100 includes: S101. Obtain a multi-angle image sequence of the clothing to be identified under natural light or artificial light source, including front, back and side views; S102. Perform image preprocessing on the image sequence; S103. Input the preprocessed image sequence into the image recognition model M0, extract the corresponding local texture features, global contour features and color histogram features to form the initial recognition feature vector F0.

3. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S110 includes: S111. Based on the color distribution and texture gradient information contained in the initial recognition feature vector F0, the superpixel clustering algorithm is used to perform preliminary segmentation of the input image to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions. S112. The candidate region is optimized by using a graph cut-based image segmentation method. S113. For each optimized candidate clothing region, extract its gray-level co-occurrence matrix to construct local texture features Lt, and use the Canny algorithm to extract edge contour features Et. S114. Jointly encode the local texture features Lt and edge contour features Et of each candidate region.

4. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S120 includes: S121. Perform dimensional unification and normalization on the local texture features and edge contour features of each candidate region to construct a standardized input feature vector; S122. Input the standardized input feature vector into the pre-trained material discrimination sub-model M1. The material discrimination sub-model M1 is constructed based on the support vector machine algorithm, and the radial basis kernel function is used to optimize the parameters through cross-validation. S123. Perform maximum value judgment on the probability distribution results of each category output by the material discrimination sub-model M1 to obtain the corresponding material discrimination label; S124. Bind the material discrimination label to the candidate area index to generate a one-to-one correspondence between candidate clothing areas and material categories.

5. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S130 includes: S131. Perform one-hot encoding on the material tag corresponding to each candidate clothing area to generate a fixed-dimensional material vector group. S132. Concatenate and splice the initial identification feature vector F0 and the material vector group in the order of regions to form a candidate region-level fusion feature matrix. S133. Use principal component analysis algorithm to perform feature dimensionality reduction on the fused feature matrix and retain the principal components with a cumulative contribution rate greater than 95%. S134. Perform global average pooling on the dimensionality-reduced fusion feature matrix to generate the fusion feature vector Ff.

6. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S140 includes: S141. Input the fused feature vector Ff into the multi-layer attention-aware network of the image semantic enhancement model M2 to extract the initial semantic context features; S142. Utilize a bidirectional feature residual connection structure to perform deep iterative optimization of the initial semantic context features; S143. Adjust the weights of each feature dimension to increase the proportion of key semantic features in the output vector; S144. The optimized semantic context features are normalized and activated to output the semantic enhancement recognition feature vector Fe.

7. The garment recognition method for an intelligent garment care machine according to claim 1, characterized in that: The S150 mentioned above includes: S151. Input the semantically enhanced recognition feature vector into the dual-channel classification structure of the clothing type and nursing needs joint recognition model, and extract clothing category features and nursing feature sub-vectors respectively. S152. The clothing category feature sub-vectors are processed by a fully connected layer and a Softmax function to output the classification result of the target clothing. S153. Input the nursing feature sub-vectors into the multi-label output layer, and combine them with the multilayer perceptron to calculate the fit probability of each nursing option. S154. Based on the classification results and the matching probability, select the nursing option with the highest probability as the recommended nursing mode, and combine it with the clothing category results for output.

8. A garment recognition system for an intelligent garment care machine, used to implement the garment recognition method for an intelligent garment care machine as described in any one of claims 1-7, characterized in that: include: The image acquisition module acquires the image input data of the clothing to be identified and inputs it into the preset image recognition model M0 to obtain the initial recognition feature vector F0. The feature extraction module performs region segmentation on the input image based on the initial recognition feature vector F0 to obtain multiple candidate clothing regions R1, R2, ..., Rn, where n is the total number of candidate clothing regions, and extracts the local texture features Lt and edge contour features Et of each region respectively. The material recognition module inputs the local texture features Lt and edge contour features Et of each region into the material discrimination sub-model M1, and outputs the material discrimination labels T1~Tn corresponding to each region. The feature fusion module fuses the initial recognition feature vector F0 with all material labels T1~Tn to generate a fused feature vector Ff; The semantic enhancement processing module iteratively optimizes the fused feature vector Ff through the image semantic enhancement model M2 to obtain the semantic enhancement recognition feature vector Fe. The joint identification module inputs the semantically enhanced identification feature vector Fe into the clothing type and nursing needs joint identification model M3, and outputs the classification result of the target clothing and the recommended nursing mode. The nursing task execution module executes the corresponding nursing tasks through the intelligent clothing care machine based on the classification results and recommended nursing modes.

Images