A method and system for intelligent sorting of clothing items for combination orders

Abstract

The present application relates to the technical field of clothing commodity sorting, in particular to a clothing commodity intelligent sorting method and system for combined orders. The method collects clothing commodity images on a conveying mechanism and pre-processes to obtain to-be-matched images, synchronously obtains current orders containing multiple commodity attributes and user identification; calls historical order data of the user, inputs a pre-trained prediction model to obtain predicted orders and confidence, compares the current and predicted orders to generate attention degrees of each commodity attribute; generates a key template according to the attention degrees, adjusts the matching threshold combined with the confidence, and generates a basic template of the order commodity; uses the key template and the set threshold to coarsely match the images, obtains a candidate region and an initial matching score; updates the score through fine matching of the candidate region by the basic template, and completes the order commodity correlation matching combined with the spatial position relationship of the region; controls a clamping unit to complete commodity gripping and packaging sorting according to the correlation result, thereby significantly reducing the probability of mis-sorting and missing sorting.

Description

Technical Field

[0001] This invention relates to the field of garment sorting technology, specifically to an intelligent sorting method and system for garments in combination orders. Background Technology

[0002] In e-commerce multi-item order sorting scenarios, the online clothing categories are complex, and the colors or sizes of the same style of clothing are very similar. Traditional sorting relies on manual differentiation, which results in low sorting efficiency and a high rate of missorting and missed sorting.

[0003] Currently, for the industry's demand for bundled packaging where multiple customers each purchase at least two garments, existing technologies still primarily rely on manual sorting and packaging. Some companies have introduced semi-automated equipment that only achieves one-way transport of garments, lacking precise product identification, customer categorization, and targeted bundling modules. This still requires significant manual intervention to match and categorize products with orders. Because garments in multiple orders on conveyor belts often have high visual similarity, relying solely on real-time order retrieval can lead to positioning ambiguity and an inability to automatically filter for highly identifiable attributes, resulting in low sorting efficiency. Furthermore, since garments encompass multiple attributes such as color, pattern, category, and size, and different users have varying consumption preferences, relying solely on single, shallow features for coarse matching to determine sorting targets is insufficient for adapting to various order types. This leads to difficulties in accurately binding products to bundled orders when multiple garments are stacked or mixed, ultimately resulting in insufficient identification accuracy. Summary of the Invention

[0004] To address the technical problems of low sorting efficiency and high error rates in existing garment sorting solutions, such as the inability to adaptively adjust to scenarios where multiple customers purchase multiple garments and need to combine and pack them, resulting in manual bundling, this invention provides an intelligent garment sorting method and system for combined orders. The specific technical solution adopted is as follows: This invention proposes an intelligent sorting method for apparel goods oriented towards combined orders, the method comprising: Images of clothing items on the conveyor are collected and preprocessed to generate images to be matched. At the same time, the current order and the corresponding user identifier are obtained. The current order contains multiple attributes of each clothing item. Retrieve the user's historical order data based on the user identifier, input the historical order data into the pre-trained prediction model to obtain the predicted order and the corresponding confidence score; compare the current order with the predicted order to generate the attention score of each attribute. Based on the attention of each attribute, at least one key attribute is determined, and the corresponding feature descriptor is read from the pre-built feature library based on the key attribute to generate a key template. The matching threshold of the key template is adjusted according to the confidence level; a basic template is generated for each product in the current order. Based on the type of key attributes, key templates are used and matching filters are performed in the image to be matched according to the matching threshold to obtain candidate regions and matching scores for each candidate region. Based on the basic template of each product, each candidate region is finely matched, the matching score of each candidate region is updated, and the spatial relationship between each candidate region is combined to determine the association result between each candidate region and each product in the current order; based on the association result, the corresponding clamping unit is controlled to perform gripping and packaging actions.

[0005] Furthermore, the process of generating the image to be matched includes: Obtain a background reference image of the conveyor mechanism under no-load conditions; Background subtraction technology is used to perform pixel-by-pixel difference processing between the acquired image and the background reference image to generate a difference image; The difference image is segmented using a preset separation threshold to extract candidate product regions from the acquired image; The candidate regions for the products are deblurred to obtain the image to be matched.

[0006] Furthermore, the prediction model is a sequence prediction model pre-trained using historical order data from multiple users; The process of obtaining the predicted order and its corresponding confidence level includes: Based on the user identifier, retrieve the user's historical order data, which includes the attributes of each clothing item purchased by the user in the past; among which, the attributes include at least category, color, size and pattern; Historical order data is input into a pre-trained prediction model. The prediction model outputs predicted orders and their corresponding confidence scores. Each predicted order contains at least one predicted product, and each predicted product has corresponding attributes.

[0007] Furthermore, the process for determining the attention level of each attribute includes: The system pairs each item in the current order with each predicted item in the predicted order to obtain paired items. Each paired item is then compared for differences in each attribute to generate an attribute difference score. Specifically, for category, color, and pattern type attributes, if the attribute value of the item in the current order is the same as its attribute value in the predicted order, the attribute difference score is zero; otherwise, it is set to one. For size-ordered numerical attributes, sizes are mapped to numerical values ​​in ascending order to obtain size values. The difference between the largest and smallest size values ​​is calculated as the maximum size difference. The absolute value of the difference between the size value of the item in the current order and the size value in the predicted order is calculated and then divided by the maximum size difference score to obtain the attribute difference score for that size attribute. Assign a preset base weight to each attribute; From the user's historical order data, we statistically analyze the value distribution of each attribute of the clothing items purchased by the user in the past; for category, color, and pattern type attributes, we use information entropy to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute; for size ordered numerical attributes, we use variance to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute. For each attribute, calculate the difference between the value one and the purchase stability of that attribute to obtain the instability of that attribute; multiply the preset modulation strength coefficient and the instability as the modulation amount; subtract the modulation amount from the value one to obtain the modulation coefficient; multiply the preset base weight of that attribute by the corresponding modulation coefficient to obtain the attribute weight of that attribute. Calculate the product of the attribute weight and the corresponding attribute difference degree for each attribute, and use it as the weighted difference degree of each product on each attribute; sum the weighted difference degrees of the same attribute on each product to obtain the total attribute difference value of each attribute; sum and normalize the total attribute difference value of each attribute to obtain the attention degree of each attribute.

[0008] Furthermore, the pre-built feature library contains feature descriptors for each clothing item, with each clothing item corresponding to a set of feature descriptors. The feature descriptors include at least the color feature descriptor, texture feature descriptor, and depth feature descriptor for the clothing item. Specifically, the color feature descriptor indicates the hue range, saturation range, and lightness range of the clothing item in the HSV color space; the texture feature descriptor indicates the local texture pattern distribution histogram of the clothing item; and the depth feature descriptor indicates the depth feature vector of the clothing item. The process of determining the matching threshold of the key template includes: Obtain the preset baseline threshold and preset relaxation coefficient; Based on the level of attention of each attribute, select at least one attribute with the highest level of attention as the key attribute; Based on the key attributes, the feature descriptors of the corresponding products are read from the pre-built feature library to generate key templates; if the key attributes contain two attributes, then two corresponding key templates are generated respectively. The difference between the numerical value and the confidence level is used as the unreliability level; the product of the preset relaxation coefficient and the unreliability level is used as the relaxation amount. The difference between the first value and the relaxation amount is calculated to obtain the adjustment factor; the product of the preset benchmark threshold and the adjustment factor is calculated as the adjusted matching threshold.

[0009] Furthermore, the process of obtaining the candidate regions and the matching scores of each candidate region includes: The image to be matched is divided into multiple matching units, each of which is a pixel or a sliding window region; If the key attribute is color, then the matching unit is pixel, and the corresponding key template is recorded as the color template; extract the color feature descriptor containing the target color in the color template; perform color matching between each pixel in the image to be matched and the color template, calculate the membership degree of each pixel to the target color, and use it as the matching response result of each pixel; If the key attribute is a pattern, the matching unit is a sliding window region, and the corresponding key template is recorded as a texture template. The image to be matched is traversed in a sliding window manner, and the texture feature descriptors of each window region are extracted in turn. The texture feature matching technology is used to calculate the chi-square distance between the texture feature descriptors of each window region and the texture template, and the chi-square distance is converted into similarity as the matching response result of each window region. The matching response results are filtered according to the matching threshold. The positions that meet the preset conditions after filtering are divided into candidate regions. The matching score of each candidate region is determined by combining the matching response results.

[0010] Further, the step of filtering the matching response results according to the matching threshold, dividing the filtered positions that meet the preset conditions into candidate regions, and determining the matching score of each candidate region based on the matching response results includes: Matching units whose matching response results reach the matching threshold are marked as valid units; adjacent valid units are connected and aggregated to form connected regions; the circumscribed regions of each connected region are determined as candidate regions; for each candidate region, the arithmetic mean of the matching response results of each matching unit contained in the candidate region on the same key attribute is calculated as the matching component of the candidate region on the same key attribute. If a candidate region has both color matching results and texture matching results, the matching components of all key attributes contained in the candidate region are weighted and fused according to the preset first weight to generate the matching score of the candidate region.

[0011] Furthermore, the basic template is a corresponding deep feature descriptor read from a pre-built feature library based on the product identifier carried by each product in the current order; The process of performing fine-grained matching on each candidate region based on the basic template of each product and updating the matching score of each candidate region includes: Extract candidate image patches corresponding to each candidate region from the image to be matched; Each candidate image patch is finely matched with the base template of each product in the current order, and the depth feature descriptor of each candidate image patch is extracted. The cosine distance between the depth feature descriptor of each candidate image patch and the depth feature descriptor of each base template is calculated, and the cosine distance is mapped to a value between zero and one, which is used as the fine matching score between each candidate image patch and each base template; where the cosine distance and the fine matching score are negatively correlated. For each candidate region, the highest score is selected from the fine matching scores between the candidate region and each product, and this score is used as the updated matching score for the candidate region. The product corresponding to the highest score and its product identifier are recorded. If the highest score is lower than the preset fine matching threshold, the candidate region is determined not to belong to any product in the current order and is removed.

[0012] Furthermore, the process for determining the association result includes: Determine the center point of each candidate region; calculate the pairwise distance between the center points of each candidate region; calculate the arithmetic mean of all pairwise distances as the average spacing between candidate regions. The ratio of the average spacing to the preset reference spacing is calculated as the spacing ratio; the spacing ratio is mapped to a value between zero and one using an exponential decay method to obtain the location clustering degree. Calculate the arithmetic mean of the updated matching scores for all candidate regions as the average refinement score; The average fine score and the location clustering degree are each assigned a preset second weight, and then weighted and fused to obtain the comprehensive confidence score. If the overall confidence level reaches or exceeds the preset decision threshold, each candidate region is determined to be successfully associated with the current order, and the association result corresponding to each candidate region is output. The association result is used to indicate whether each candidate region belongs to the current order and which product in the current order it belongs to. If the overall confidence level is lower than the preset decision threshold, the association is deemed to have failed, and the process is transferred to the exception handling procedure.

[0013] A smart sorting system for apparel goods oriented to combined orders, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of a smart sorting method for apparel goods oriented to combined orders.

[0014] The present invention has the following beneficial effects: This invention retrieves users' historical order data and uses a pre-trained prediction model to generate predicted orders and confidence levels. It then compares the current order with the predicted order to generate attention levels for each attribute, effectively solving the problem of ambiguous sorting targets caused by relying solely on real-time order information. This makes it suitable for sorting scenarios with numerous clothing categories and highly similar styles. Based on the confidence level predicted from historical order data, the invention adaptively adjusts the key template matching threshold. When the confidence level is low, the relaxation of the matching threshold is increased to avoid missed detections; when the confidence level is high, the matching threshold is tightened to filter out invalid interference areas, improving the sorting target detection rate and overcoming the poor adaptability of fixed matching thresholds. First, key attribute templates are used to perform coarse matching to screen candidate areas. Then, basic templates specific to each product are used for fine matching to update the matching score. This achieves multi-layer feature fusion to improve product recognition accuracy. Furthermore, it introduces spatial location clustering constraints of candidate areas to comprehensively determine the association results, thereby achieving accurate classification of each product in scenarios with multiple stacked and mixed garments. This significantly reduces the probability of mis-picking and missed picking, greatly improving the automation and efficiency of sorting combined orders. Attached Figure Description

[0015] Figure 1 A flowchart illustrating an intelligent sorting method for apparel goods oriented towards combined orders, provided in one embodiment of the present invention; Figure 2 This is an example diagram of an intelligent sorting system for apparel goods oriented towards combined orders, provided in one embodiment of the present invention. Figure 3 This is an example diagram illustrating the process of determining the attention level of various attributes provided in one embodiment of the present invention.

[0016] Figure 2 The labels in the text are denoted as: 1. Garment goods to be sorted; 2. Conveyor belt; 3. Rotating roller; 4. Camera; 5. Rotary disk; 6. Clamping device; 7. Storage device; 8. Packaging bag; 9. Control system. Detailed Implementation

[0017] The following description, in conjunction with the accompanying drawings, details a specific solution for an intelligent sorting method and system for combined orders of clothing provided by this invention.

[0018] Please see Figure 1 The diagram illustrates a flowchart of an intelligent sorting method for apparel goods oriented towards combined orders, according to an embodiment of the present invention. The method includes: S101: Collect and preprocess images of clothing items on the conveyor mechanism to generate images to be matched. At the same time, obtain the current order and the corresponding user identifier. The current order contains multiple attributes of each clothing item.

[0019] A conveyor mechanism is a mechanical device that carries and transports clothing merchandise, such as a conveyor belt. Specifically, clothing merchandise is placed on the conveyor mechanism and passes sequentially through an image acquisition and clamping unit.

[0020] For example, an example diagram of an intelligent sorting system for apparel goods oriented towards combined orders is shown below. Figure 2 As shown, 1 represents clothing items to be sorted, each covered with an individual transparent package; 2 represents a conveyor belt used to transport the clothing items; 3 represents rotating rollers used to power the conveyor belt 2, and the rotating rollers 3 are usually configured in pairs to ensure smooth transport by the conveyor belt 2; 4 represents a camera used to capture images of clothing items 1 and transmit the images in real time to the control system 9 for storage, providing a data basis for subsequent product classification; 5 represents a rotating disk that can rotate under the control of the control system 9, thereby adjusting the spatial position of the clamping device 6; 6 represents the clamping device, with multiple clamping devices evenly arranged circumferentially below the rotating disk 5, each of which can be independently controlled for use in... The system performs telescopic and gripping / releasing actions, with each gripping device 6 gripping one garment item 1. 7 represents a storage device, located at the unloading station of the rotary table 5, equipped with a ramp structure. When the gripping device 6 releases the garment item 1, it falls into the storage device 7. 8 represents a packaging bag, serving as the final packaging carrier for the garment item 1. 9 represents the control system, the core control unit of the entire intelligent garment sorting system for combined orders. It establishes signal connections with the camera 4, rotating roller 3, rotary table 5, and gripping devices 6, enabling the storage of product information, allocation and control of gripping devices 6, analysis and classification of user information, and automated execution of the packaging process.

[0021] It should be noted that in this embodiment, the conveyor belt 2 and the rotating roller 3 together constitute the conveying mechanism described in this invention, used to realize the transfer of clothing goods between different workstations. Furthermore, in other feasible embodiments, the conveying mechanism can also be replaced by chain conveyors, roller conveyors, etc., all of which are common and simple substitutions in the art.

[0022] In existing industrial image acquisition scenarios, the object being photographed typically needs to face the camera and the background should be as clean as possible. Therefore, the image acquisition of clothing products in this invention also follows these conventional requirements. For example, when the clothing product passes the camera 4 on the conveyor, its product label (such as a barcode) should be placed facing upwards so that the camera 4 can capture a clear image.

[0023] In addition, at the feeding end of the conveyor, multiple garment items under the same combined order are pre-loaded into the same small area on the conveyor belt by an upstream flipping machine or a batch feeding device according to the order, so that the goods of the same order form a basic aggregation on the conveyor belt.

[0024] The current order is a combination order submitted by a user that is currently being processed.

[0025] The user identifier is a unique identifier for a user, used to retrieve the user's historical order data from the order database.

[0026] Attributes are dimensions that describe the characteristics of clothing products. The attributes used in this invention include category, color, size, and pattern, where each attribute corresponds to a specific attribute value. For example, product number 1 is a shirt, the color is blue, the size is L, and the pattern is solid color; product number 2 is a T-shirt, the color is white, the size is M, and the pattern is striped.

[0027] In this embodiment, a background reference image of the conveyor mechanism under no-load conditions is acquired; a background subtraction technique is used to perform pixel-by-pixel difference processing on the acquired image and the background reference image to generate a difference image; a preset separation threshold is used to segment the difference image and extract the candidate product region in the acquired image; the candidate product region is deblurred to obtain the image to be matched.

[0028] The conveyor system is considered unloaded when no clothing or merchandise is placed on it. In this case, the camera only captures a background reference image of the fixed background, such as the surface of the conveyor belt.

[0029] Specifically, the grayscale value of the currently captured image containing clothing is subtracted from the grayscale value of the previously captured background reference image at the same pixel. If the grayscale value of a pixel is the same or similar in both the captured image and the background reference image, it means that the location of that pixel is more likely to belong to the background (such as a conveyor belt), and the difference after subtraction is very small or zero; conversely, if the grayscale value of a pixel is different in the two images, it means that the location of that pixel is more likely to be a product on the conveyor belt.

[0030] It should be noted that the specific value of the preset separation threshold can be determined based on engineering experience, and this embodiment does not impose a specific limitation. For example, if the grayscale value is in the range of 0 to 255, the preset separation threshold is usually between 20 and 40.

[0031] It should be noted that the deblurring process can be implemented using Wiener filtering or the Lucy-Richardson algorithm, both of which are mature technologies in the field of image processing, and will not be elaborated on here.

[0032] Since the conveyor belt is constantly in motion during the sorting process, and the clothing items are also in motion when the camera takes pictures, the captured images may have motion blur (such as image edge blurring or unclear details). Therefore, after obtaining the candidate areas of the items, these areas are deblurred to make the images to be matched clear.

[0033] S102: Retrieve the user's historical order data based on the user identifier, input the historical order data into the pre-trained prediction model, obtain the predicted orders and their corresponding confidence scores; compare the current orders with the predicted orders to generate the attention scores of each attribute.

[0034] Since e-commerce platforms typically store a complete record of a user's past purchases, including fields such as order number, product number, category, color, size, and pattern, historical order data can be retrieved using a standard database query based on the user's identifier.

[0035] The confidence score is a numerical value between 0 and 1 output by the prediction model, used to represent the reliability of the prediction result (i.e., the predicted order). The more reliable the predicted order, the higher the confidence score; the less reliable the predicted order, the lower the confidence score.

[0036] It should be noted that the pre-trained prediction model can be a sequence prediction model (such as Transformer), trained offline before the system goes live. Specifically, during training, a large amount of historical order data from users is extracted; historical order data of the same user sorted by time is used as input, and the actual product combinations purchased by the user in the corresponding time period are used as labels. The model is trained using supervised learning, enabling it to predict the most likely product combinations a user will purchase based on their historical purchasing behavior, ultimately yielding the prediction model. The above model structure and training methods are mature technologies in the field of machine learning, and those skilled in the art can adjust them according to the actual data scale; further details are omitted here.

[0037] In this embodiment, the user's historical order data is obtained based on the user's identifier. The historical order data includes the attributes of each clothing item purchased by the user in the past. The attributes include at least category, color, size, and pattern. The historical order data is input into a pre-trained prediction model, and the prediction model outputs predicted orders and corresponding confidence scores. The predicted orders include at least one predicted item, and each predicted item has corresponding attributes.

[0038] It's important to understand that clothing products possess multiple attributes such as category, color, size, and pattern. The visual salience of these attributes varies significantly in images. For example, color is the most intuitive and stable, while patterns are greatly affected by deformation. Category and size are more difficult to directly identify in images. Therefore, treating all attributes equally without distinction would lack prioritization in visual sorting, leading to inefficiency. Thus, by assessing the actual differences between current and predicted orders across attributes, combined with the stability of a user's historical purchase records for each attribute, the focus on each attribute can be evaluated. This allows for determining which attribute to prioritize in subsequent steps, providing a reliable data foundation for visual sorting.

[0039] The process of determining the attention level of each attribute is as follows: Figure 3 As shown, it includes: S201: Pair each item in the current order with each predicted item in the predicted order to obtain paired items; perform a difference comparison on each attribute of each paired item to generate the attribute difference degree of each item; for category, color, and pattern type attributes, if the attribute value of the item in the current order is the same as the attribute value in the predicted order, the attribute difference degree is zero; otherwise, if they are different, the attribute difference degree is set to one; for size ordered numerical attributes, map the sizes to numerical values ​​in ascending order to obtain size values, and calculate the difference between the largest and smallest size values ​​as the maximum size difference value; calculate the absolute value of the difference between the size value of the item in the current order and the size value in the predicted order, and then divide it by the maximum size difference value to obtain the attribute difference degree of the size attribute.

[0040] It should be noted that assignment algorithms from combinatorial optimization, such as the Hungarian algorithm, can be used. This approach treats any item and any predicted item as a combination. After iterating through all possible combinations, the combination that maximizes the number of identical attributes is selected as the final pair of items. These assignment algorithms are mature techniques in combinatorial optimization and will not be elaborated upon further here.

[0041] It should be noted that the paired products have the most identical attributes overall.

[0042] For example, the current order includes: Product A is a shirt (blue, size L, solid color), and Product B is a T-shirt (white, size M, striped); the predicted order includes: predicted Product C is a shirt (blue, size M, solid color), and predicted Product D is a T-shirt (blue, size L, striped). The optimal pairing is the pairing of products with the same category and the closest other attributes. So, the shirt in the current order is paired with the shirt in the predicted order, i.e., Product A is paired with the predicted Product C; the T-shirt in the current order is paired with the T-shirt in the predicted order, i.e., Product B is paired with the predicted Product D.

[0043] For example, if the current order is blue and the predicted order is blue, the attribute difference is 0; if the current order is striped and the predicted order is solid color, the attribute difference is 1.

[0044] Attribute variance measures the degree of difference in a particular attribute between a matched pair of goods in the current order and the predicted order. A higher attribute variance indicates a greater deviation between the actual and predicted attribute values, making the attribute more worthy of attention. Conversely, a zero attribute variance indicates a closer match between the actual and predicted attribute values, requiring less attention to that attribute.

[0045] It should be noted that the sizes are mapped to consecutive numerical values ​​in ascending order. Specifically, size S→1, size M→2, size L→3, size XL→4, size XXL→5.

[0046] The maximum size difference is used to normalize the absolute value of the difference between size values ​​so that the range of values ​​for the size attribute's difference is consistent with other attributes such as category, color, and pattern.

[0047] S202: Assign preset base weights to each attribute.

[0048] It should be noted that different attributes are assigned different preset base weights to reflect the visual salience of each attribute in the image.

[0049] Visual saliency refers to the degree of difficulty with which a certain attribute can be identified and distinguished in an image.

[0050] Color is generally the most easily perceived attribute and remains relatively stable even under different lighting conditions; patterns are more affected by clothing deformation and wrinkles, making them slightly more difficult to identify; categories require recognizing the overall outline of the garment and are greatly affected by occlusion; sizes lack absolute reference points in images, making them difficult to judge directly and accurately based solely on the image. Therefore, color has the highest visual salience, followed by patterns, while categories and sizes have relatively lower salience.

[0051] It should be noted that the specific values ​​of the preset basic weights for each attribute can be adjusted based on the visual saliency differences of each attribute in the image, and this embodiment does not impose specific limitations. For example, the sum of the four preset basic weights—color 0.5, pattern 0.3, category 0.1, and size 0.1—is 1.

[0052] S203: From the user's historical order data, statistically analyze the value distribution of each historical clothing item purchased by the user in each attribute; for category, color, and pattern type attributes, use information entropy to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute; for size ordered numerical type attributes, use variance to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute.

[0053] Information entropy is used to quantify the concentration of values ​​for a certain attribute. Specifically, it involves counting the frequency of each value for a certain attribute for all historical clothing items purchased by a user, multiplying these frequencies by their logarithms, summing them, and then inverting the sum to obtain the information entropy. The information entropy is then divided by the logarithm of the total number of possible values ​​for that attribute to achieve normalization, resulting in a normalized information entropy that falls between 0 and 1. Finally, the purchase stability is calculated as 1 - normalized information entropy.

[0054] It should be noted that in the actual clothing product system, the categories (shirts, T-shirts, pants, etc.), colors (red, blue, black, etc.), and patterns (solid colors, stripes, checks, etc.) all naturally have a number of possible values ​​that are much greater than 2. Therefore, the logarithm of the total number of possible values ​​is a positive number, meaning that the logarithm of the total number of possible values ​​cannot be zero.

[0055] For example, if a user has purchased 20 clothing items in the past, 15 are blue, 3 are white, and 2 are black, the color distribution is very concentrated, the information entropy is low, and the purchase stability is high; if the color distribution is 4 blue, 5 white, 6 black, and 5 red, the color distribution is scattered, the information entropy is high, and the purchase stability is low.

[0056] Since sizes have been mapped to numerical values ​​in ascending order, the variance of all historical clothing purchases by a user is calculated. This variance is then normalized by dividing by a pre-defined theoretical maximum variance, yielding the normalized variance. Finally, purchase stability is calculated as 1 - normalized variance. A smaller normalized variance indicates that size values ​​are more concentrated around a specific size range, resulting in higher purchase stability; conversely, a larger normalized variance indicates that size values ​​are more dispersed, leading to lower purchase stability.

[0057] It should be noted that the specific value of the preset theoretical maximum variance can be obtained through statistical analysis, and this embodiment does not impose a specific limitation. For example, a large amount of historical order data can be obtained; the variance values ​​when all possible values ​​of each attribute are evenly distributed can be calculated, and the median of all possible variance values ​​of each attribute can be taken as the preset theoretical maximum variance. For example, if there are 5 size values ​​(such as S, M, L, XL, XXL mapped to 1, 2, 3, 4, 5), and these values ​​are evenly distributed (such as each accounting for 20%), the expected value is 3, and the variance is 2, then the preset theoretical maximum variance is 2.

[0058] S204: For each attribute, calculate the difference between the value one and the purchase stability of the attribute to obtain the instability of the attribute; multiply the preset modulation strength coefficient and the instability as the modulation amount; subtract the modulation amount from the value one to obtain the modulation coefficient; multiply the preset basic weight of the attribute by the corresponding modulation coefficient to obtain the attribute weight of the attribute.

[0059] It should be noted that the specific value of the preset modulation intensity coefficient can be determined through experimental calibration, and this embodiment does not impose a specific limitation. A specific experimental process is as follows: Extract a certain amount of historical order data, and for different preset modulation intensity coefficient values, calculate the attention level of each attribute in each historical order data set. Compare the attribute with the highest attention level under different preset modulation intensity coefficient values ​​with the manually labeled "most attentionable" attributes in the historical order data. Finally, find the value that results in the highest consistency rate as the optimal preset modulation intensity coefficient. For example, a typical range for the preset modulation intensity coefficient is 0.5 to 0.8.

[0060] If a user's purchase stability on a certain attribute is high (i.e., purchase stability approaches 1), it indicates that the user's purchasing behavior has strong regularity on that attribute, and the prediction result of the order on that attribute is more reliable. At this time, the instability of that attribute approaches 0, the modulation amount is zero, the modulation coefficient approaches 1, and the preset basic weight remains unchanged. Conversely, if a user's purchase stability on that attribute is low (i.e., purchase stability approaches 0), it indicates that the user's purchasing behavior on that attribute is volatile, and the prediction result of the order on that attribute is less reliable. At this time, the instability of that attribute approaches 1, the modulation amount approaches the preset modulation strength coefficient (i.e., 0.5 to 0.8), then the modulation coefficient approaches 0.2 to 0.5, and the preset basic weight is significantly weakened.

[0061] S205: Calculate the product of the attribute weight and the corresponding attribute difference for each attribute, and use it as the weighted difference for each product on each attribute; sum the weighted differences for the same attribute on each product to obtain the total attribute difference for each attribute; sum and normalize the total attribute difference for each attribute to obtain the attention level for each attribute.

[0062] An example of the process of summing and normalizing the total difference value of each attribute is as follows: divide the total difference value of each attribute by the sum of the total differences values ​​of all attributes, so that the sum of the attention of each attribute is one.

[0063] It should be noted that in actual sorting scenarios, the probability of the current order being exactly the same as the predicted order is extremely low, and the sum of the differences cannot be zero.

[0064] Attention level is used to measure the importance of each attribute in the current order. A higher attention level for a particular attribute indicates that the attribute is more worthy of priority in subsequent visual searches within the current order; a lower attention level indicates that the attribute's differences are not significant enough or its predictions are not reliable enough.

[0065] It's important to note that attention level determines which attribute is prioritized for matching, rather than forcing it to be used. Therefore, during the actual matching process, if the image region corresponding to a key attribute suffers from poor feature extraction quality due to lighting or occlusion, its matching score will naturally be low, and the system can automatically revert to matching the next highest attention attribute. In summary, the role of attention level is to establish a logical connection between significant differences and matching priority, rather than replacing image quality assessment; the impact of image quality is reflected in the subsequent calculation of the matching score, and both work together at different levels to complete the sorting and recognition.

[0066] S103: Determine at least one key attribute based on the attention of each attribute, and read the corresponding feature descriptor from the pre-built feature library based on the key attribute to generate a key template, and adjust the matching threshold of the key template according to the confidence level; generate basic templates for each product in the current order.

[0067] It should be noted that one or two of the most popular attributes are selected as key attributes. Specifically, if the difference between the most popular attribute and the second most popular attribute is less than a preset difference threshold, then both the most popular attribute and the second most popular attribute are selected as key attributes.

[0068] It should be noted that the specific value of the preset difference threshold can be determined through experimental calibration. A specific experimental process is as follows: During the system development phase, using a large amount of historical order data, test the rationality of selecting key attributes under different difference thresholds. This avoids the tendency to select two attributes if the difference threshold is too small, or to select only one attribute if the difference threshold is too large, preventing the possibility of missing less important attributes. Select a difference threshold value that ensures recognition accuracy while having a moderate computational load as the fixed preset difference threshold for the system. For example, the preset difference threshold value could be 0.1.

[0069] In actual sorting scenarios, the attribute that receives the most attention is color, and the attribute that receives the second most attention is pattern.

[0070] It is important to note that category attributes have the lowest visual saliency in images, requiring complete contour information for identification. They are susceptible to occlusion and deformation interference, and the category purchasing preferences of the same user are typically highly stable, with very little significant deviation in category differences. Size attributes lack absolute reference points in monocular 2D images, making reliable measurement impossible through a single viewpoint. Furthermore, the size preferences of the same user are highly stable, and their attention is further weakened after stability modulation, making them extremely rare as key attributes. Based on these facts, in actual garment sorting scenarios, the selection of category and size as key attributes is extremely rare. This embodiment only requires configuring independent visual initial screening matching methods for color and pattern, which is sufficient to cover all engineering requirements in actual sorting scenarios. Moreover, if, in extreme cases, category or size is selected as a key attribute, the system will automatically revert to the default process, using color as the key attribute for initial screening, to ensure process integrity.

[0071] The pre-built feature library contains feature descriptors for each clothing item. Each clothing item corresponds to a set of feature descriptors, which include at least the color feature descriptor, texture feature descriptor, and depth feature descriptor for that clothing item. The color feature descriptor indicates the hue range, saturation range, and lightness range of the clothing item in the HSV color space. The texture feature descriptor indicates the local texture pattern distribution histogram of the clothing item. The depth feature descriptor indicates the depth feature vector of the clothing item.

[0072] Color feature descriptors are a set of numerical ranges. For example, the hue (H) of a blue shirt ranges from 200° to 230°, the saturation (S) ranges from 0.6 to 0.9, and the lightness (V) ranges from 0.4 to 0.7.

[0073] A texture feature descriptor is a histogram vector. For example, a 59-dimensional array, where each dimension represents the frequency of a certain local texture pattern.

[0074] A deep feature descriptor is a fixed-length floating-point vector that can be extracted from an image by an existing neural network to represent the deep semantic features of a particular clothing item. For example, a deep feature descriptor consists of 128 floating-point numbers, such as [0.23, -0.45, 0.67, ...].

[0075] It should be noted that color feature descriptors can be obtained by converting the preprocessed image of the product to the HSV color space and then statistically analyzing the numerical range of each channel in the main color region; texture feature descriptors can be obtained by extracting local binary pattern (e.g., LBP) histograms from the deblurred candidate regions of the product; and depth feature descriptors can be obtained by inputting the preprocessed image into a lightweight convolutional neural network (e.g., MobileNetV3) and extracting the output vector of its penultimate layer. These feature extraction methods are all mature technologies in the field of computer vision and are performed offline during product warehousing; therefore, they will not be elaborated upon further here.

[0076] SKU (Stock Keeping Unit) is used in the apparel retail and e-commerce sectors to uniquely identify a specific product and can serve as a product identifier.

[0077] It should be noted that in the pre-built feature library, each SKU corresponds to a set of independent feature descriptors (including color feature descriptors, texture feature descriptors, and depth feature descriptors). During sorting, the system reads the feature descriptor corresponding to the product from the pre-built feature library based on the product identifier (i.e., SKU).

[0078] In one preferred implementation, a preset benchmark threshold and a preset relaxation coefficient are obtained; based on the attention of each attribute, at least one attribute with the highest attention is selected as a key attribute; based on the key attribute, the feature descriptor of the corresponding product is read from the pre-built feature library to generate a key template; wherein, if the key attribute contains two attributes, two corresponding key templates are generated respectively; the difference between the first value and the confidence level is calculated as the unreliability level; the product of the preset relaxation coefficient and the unreliability level is calculated as the relaxation amount; the difference between the first value and the relaxation amount is calculated to obtain the adjustment factor; the product of the preset benchmark threshold and the adjustment factor is calculated as the adjusted matching threshold.

[0079] It should be noted that the specific value of the preset benchmark threshold is related to the attribute type and can be determined through experimental calibration; this embodiment does not impose specific limitations. A specific experimental process is as follows: extract images of a batch of products with correctly matched results, run the matching algorithm under different benchmark thresholds; calculate the precision and recall (these are well-known evaluation metrics in the art, and their calculation methods can be directly applied); select the harmonic mean of precision and recall as the preset benchmark threshold. For example, a typical preset benchmark threshold for color matching is 0.6; a typical preset benchmark threshold for texture matching is 0.55.

[0080] It should be noted that the specific value of the preset relaxation coefficient can be determined according to actual needs, and this embodiment does not impose a specific limitation. For example, the preset relaxation coefficient can be 0.3.

[0081] It should be noted that if the key attribute is color, the color feature descriptor of the corresponding product in the current order is read from the feature library to generate a color template; if the key attribute is pattern, the texture feature descriptor of the corresponding product in the current order is read from the feature library to generate a texture template.

[0082] It can be understood that the color feature descriptor read from the pre-built feature library is the color template, and the texture feature descriptor is the texture template.

[0083] A matching threshold is a boundary used to determine whether a match is made or not. A higher matching threshold means stricter matching requirements, in which case only highly similar items are considered a match; a lower matching threshold means more lenient matching requirements, in which case partial similarity is considered a match.

[0084] It should be noted that the matching threshold mentioned in this invention is not a physical optical parameter of the camera or sensor, but a threshold for the candidate region screening process. It is a parameter at the algorithm decision level. Setting this threshold is used to determine which regions can enter the subsequent fine matching process. Its adjustment is based on the confidence of the prediction model. The influence of environmental interference factors (such as motion blur) has been eliminated by deblurring in the previous preprocessing step. Therefore, the image quality entering the matching process has been normalized to a stable level, and the matching threshold does not need to be adjusted for physical interference.

[0085] Since confidence level reflects the reliability of predicted orders, a higher confidence level (i.e., a more reliable predicted order) allows for a higher matching threshold to ensure accurate matching results and reduce false positives. Conversely, a lower confidence level (i.e., a less reliable predicted order) requires a more relaxed matching threshold to prevent targets from being missed due to an overly strict threshold.

[0086] It's important to note that adjusting the confidence level for the matching threshold doesn't directly map user behavior parameters to optical physical parameters. Instead, it treats the matching threshold as a checkpoint in the candidate region selection process, rather than a physical parameter of the camera itself. Specifically, the level of this checkpoint determines the system's tolerance for candidate regions; a high checkpoint only accepts highly similar regions, while a low threshold accepts some similar regions.

[0087] As an example, based on the product identifier of each product in the current order, the corresponding deep feature descriptor (i.e., the deep feature vector of the product) is read from the pre-built feature library and used as the basic template for each product.

[0088] S104: Based on the type of key attribute, use a key template and perform matching filtering in the image to be matched according to the matching threshold to obtain candidate regions and matching scores for each candidate region.

[0089] It's important to understand that because clothing items move continuously on the conveyor belt, the image captured by the camera may contain multiple items, with the current order's item being just one of them. Therefore, directly matching the entire image against the template would be computationally expensive and easily affected by background and other items. A better approach is to perform a quick initial screening within the image, using the key template to identify all regions resembling the sorting target as candidate regions. A matching score is then calculated for each candidate region, representing its degree of match with the key template. Subsequent steps only require fine-tuning these candidate regions, eliminating the need to process the entire image.

[0090] The process of obtaining candidate regions and their matching scores includes the following steps: 1) Divide the image to be matched into multiple matching units, each of which is a pixel or a sliding window region.

[0091] The matching unit is the smallest unit of the matching operation. If the key attribute is color, the matching unit is a pixel, and the key template is denoted as the color template; if the key attribute is pattern, the matching unit is a sliding window region, and the key template is denoted as the texture template. Then, the image to be matched is traversed in a sliding window manner. At each position, the texture feature descriptor (i.e., LBP histogram) of that window region is extracted.

[0092] It should be noted that the specific value of the sliding window area can be preset according to the image resolution and the actual imaging size of the goods on the conveyor belt; this embodiment does not impose specific limitations. For example, the sliding window size is usually 64×64 pixels, and the step size is usually 16 pixels.

[0093] 2) If the key attribute is color, the matching unit is pixel, and the corresponding key template is recorded as color template; extract the color feature descriptor containing the target color in the color template; perform color matching between each pixel in the image to be matched and the color template, calculate the membership degree of each pixel to the target color, and use it as the matching response result of each pixel.

[0094] The color template contains a color feature descriptor for the target color, namely the hue range, saturation range, and lightness range of the target color in the HSV color space. The target color is the color attribute value of the corresponding product in the current order, such as "blue".

[0095] In this context, it can be understood that each pixel is judged to determine whether its color is close to the color feature descriptor (i.e., hue range H, saturation range S, and lightness range) of the target color contained in the color template.

[0096] It should be noted that membership degree is a concept in fuzzy mathematics, representing the degree to which a pixel belongs to the target color. Its value ranges from 0 to 1, and it is calculated using the Gaussian membership function, which is common knowledge in this field and can be directly used here. Specifically, the closer the membership degree is to 1, the closer the pixel color is to the target color; the closer it is to 0, the greater the difference.

[0097] 3) If the key attribute is a pattern, the matching unit is a sliding window region, and the corresponding key template is recorded as a texture template. The image to be matched is traversed in a sliding window manner, and the texture feature descriptors of each window region are extracted in turn. The texture feature matching technology is used to calculate the chi-square distance between the texture feature descriptors of each window region and the texture template, and the chi-square distance is converted into similarity as the matching response result of each window region.

[0098] It should be noted that chi-square distance is a classic distance metric in image processing used to compare the differences between the distributions of two histograms. It is widely used in texture matching, and its calculation method is common knowledge, so it can be applied directly here. A smaller chi-square distance indicates that the two histogram distributions are more similar.

[0099] Specifically, the chi-square distance can be converted into similarity using an inverse proportional formula: similarity = 1 ÷ (1 + chi-square distance). If the chi-square distance is 0, it means the two histograms are completely identical, and the similarity is 1; the larger the chi-square distance, the closer the similarity is to 0.

[0100] 4) Filter the matching response results according to the matching threshold, divide the positions that meet the preset conditions after filtering into candidate regions, and determine the matching score of each candidate region in combination with the matching response results.

[0101] In a preferred embodiment, matching units whose matching response results reach the matching threshold are marked as valid units; adjacent valid units are connected and aggregated to form connected regions; the circumscribed regions of each connected region are determined as candidate regions; for each candidate region, the arithmetic mean of the matching response results of each matching unit contained in the candidate region for the same key attribute is calculated as the matching component of the candidate region for the same key attribute; if a candidate region has both color matching results and texture matching results, the matching components of all key attributes contained in the candidate region are weighted and fused according to a preset first weight to generate the matching score of the candidate region.

[0102] It should be noted that connectivity aggregation is a common operation in image processing, such as merging adjacent effective units into a single connected region (e.g., 4-connected or 8-connected). This will not be elaborated upon here.

[0103] The circumscribed region is the smallest rectangular boundary that can perfectly enclose a connected region. Specifically, for the position coordinates of all valid units within the connected region, take the minimum and maximum values ​​of the x-coordinate and the minimum and maximum values ​​of the y-coordinate, respectively. The rectangular region defined by the diagonal vertices of (minimum x-coordinate, minimum y-coordinate) and (maximum x-coordinate, maximum y-coordinate) is the circumscribed region.

[0104] It should be noted that the specific value of the preset first weight can be determined based on the importance attached to color and pattern in the actual sorting scenario, and this embodiment does not impose a specific limitation. For example, the preset first weight for the key attribute of color is 0.6; the preset first weight for the key attribute of pattern is 0.4.

[0105] Furthermore, if there is only one key attribute (such as only color or only pattern), then there is only one matching result in the candidate region, and there is no need to perform weighted fusion. The matching component is the matching score of the candidate region.

[0106] S105: Based on the basic template of each product, perform fine matching on each candidate region, update the matching score of each candidate region, and determine the association result between each candidate region and each product in the current order by combining the spatial position relationship between each candidate region; based on the association result, control the corresponding clamping unit to perform gripping and packaging actions.

[0107] It's important to understand that the candidate regions and their matching scores obtained through the initial screening using key templates only indicate that the candidate regions are similar to the sorting targets in terms of a single visual attribute such as color or pattern. They don't confirm which specific item in the current order the candidate region corresponds to. Therefore, a more refined matching can be performed on each candidate region using the basic templates (i.e., deep feature descriptors) of each item. The fine-grained matching score between each candidate region and each item can be calculated one by one, thus accurately mapping the candidate regions to specific items. The matching score is then updated accordingly, providing an accurate basis for subsequent association determination and sorting execution.

[0108] The basic template is a corresponding deep feature descriptor read from a pre-built feature library based on the product identifiers carried by each product in the current order.

[0109] In this embodiment, candidate image blocks corresponding to each candidate region are extracted from the image to be matched; each candidate image block is finely matched with the base template of each product in the current order to extract the depth feature descriptor of each candidate image block; the cosine distance between the depth feature descriptor of each candidate image block and the depth feature descriptor of each base template is calculated, and the cosine distance is mapped to a value between zero and one as the fine matching score between each candidate image block and each base template; wherein, the cosine distance and the fine matching score are negatively correlated; for each candidate region, the highest score is selected from the fine matching scores between the candidate region and each product as the updated matching score of the candidate region, and the product corresponding to the highest score and its product identifier are recorded; wherein, if the highest score is lower than the preset fine matching threshold, it is determined that the candidate region does not belong to any product in the current order and the candidate region is removed.

[0110] A candidate image patch is a sub-image cropped from the image to be matched according to the position of the candidate region. Each candidate region corresponds to one candidate image patch.

[0111] It should be noted that candidate image patches can be input into a lightweight convolutional neural network (such as MobileNetV3). The network converts the candidate image patch into a fixed-length vector through multiple convolution and pooling operations. This vector is the deep feature descriptor of the candidate image patch.

[0112] Since a smaller cosine distance reflects a smaller angle between the depth feature descriptors of each candidate image patch and the depth feature descriptors of each base template, meaning the two vectors are more aligned and the image content is more similar, there is a negative correlation between cosine distance and fine-grained matching score.

[0113] Since the cosine distance ranges from 0 to 2, a fine-match score is calculated by mapping the cosine distance to a value between 0 and 1, so that a higher fine-match score indicates greater similarity. Specifically, the fine-match score is calculated as 1 - (cosine distance ÷ 2). Wherein, if the cosine distance is 0 (i.e., the two vectors are in the same direction with an angle of 0°), the fine-match score is 1; if the cosine distance is 2 (i.e., the two vectors are in opposite directions with an angle of 180°), the fine-match score is 0; and if the cosine distance is 1 (i.e., the angle is 90°), the fine-match score is 0.5.

[0114] It's important to understand that fine-grained matching relies solely on image content similarity. However, if other visually similar items exist on the conveyor belt, the fine-grained matching score might be artificially inflated, leading to false positives. Furthermore, fine-grained matching can miss detections due to blurred or partially occluded product images, failing to capture the true physical distribution pattern of similar items clustering in space. Therefore, considering that items from the same order are fed in concentrated batches using the aforementioned method, maintaining relatively close spatial relationships during subsequent transport, the candidate regions for items from the same order exhibit clustered distributions in the image. By combining the prior knowledge that the spatial relationships of these candidate regions—that is, the proximity of items from the same order on the conveyor belt—a weighted fusion of visual similarity and spatial clustering can be used to determine the association between each candidate region and the items in the current order, effectively filtering out visual false positives.

[0115] In a preferred implementation, the center point of each candidate region is determined; the pairwise distances between the center points of each candidate region are calculated; the arithmetic mean of all pairwise distances is calculated as the average spacing between each candidate region; the ratio of the average spacing to a preset reference spacing is calculated as the spacing ratio; the spacing ratio is mapped to a value between zero and one using an exponential decay method to obtain the location clustering degree; the arithmetic mean of the updated matching scores of all candidate regions is calculated as the average refinement score; the average refinement score and the location clustering degree are each assigned a preset second weight and weighted and fused to obtain the comprehensive confidence level; if the comprehensive confidence level reaches or exceeds a preset decision threshold, each candidate region is determined to be successfully associated with the current order, and the association result corresponding to each candidate region is output; wherein, the association result is used to indicate whether each candidate region belongs to the current order and which product in the current order it belongs to; if the comprehensive confidence level is lower than the preset decision threshold, the association is determined to be unsuccessful, and the process is transferred to the exception handling procedure.

[0116] The center point is the geometric center of the candidate region. Specifically, it is obtained by calculating the center coordinates of the smallest bounding rectangle of the candidate region: the horizontal coordinate is the arithmetic mean of the left and right boundaries of the rectangle, and the vertical coordinate is the arithmetic mean of the top and bottom boundaries of the rectangle.

[0117] It can be understood that the pairwise distance between two center points is the Euclidean distance.

[0118] It should be noted that the specific value of the preset reference spacing can be determined based on the actual situation of the conveyor belt, and this embodiment does not impose a specific limitation. For example, a typical value of the preset reference spacing is one-third of the width of the conveyor belt.

[0119] Specifically, location clustering = If the average spacing is much smaller than the preset reference spacing, it indicates that the candidate regions are closely clustered, and the smaller the spacing ratio, the closer the location clustering is to 1; if the average spacing is much larger than the preset reference spacing, it indicates that the candidate regions are scattered, and the larger the spacing ratio, the closer the location clustering is to 0.

[0120] It should be noted that the specific value of the preset second weight can be determined based on engineering experience, and this embodiment does not impose a specific limitation. For example, the typical value of the preset second weight for the average refinement score and the preset second weight for the location clustering degree is 0.5.

[0121] It should be noted that the specific value of the preset decision threshold can be obtained based on statistical analysis, and this embodiment does not impose a specific limitation. Specifically, a large amount of historical order data is collected as a sample set; on the sample set, the accuracy rate under different decision thresholds is tested, that is, accuracy rate = (number of samples correctly judged under the decision threshold ÷ total number of samples) × 100%; then the accuracy rate values ​​under each decision threshold are compared, and the decision threshold with the highest accuracy rate is selected as the preset decision threshold.

[0122] Anomaly handling process: The current order and its corresponding candidate area information are sent to the manual review terminal, where staff manually review the image and the current order to complete the matching confirmation or correction.

[0123] It should be noted that, by Figure 2 As can be seen, the multiple clamping devices 6 below the rotating disk 5 in this embodiment are the clamping units described in this invention, which are used to independently clamp each garment item.

[0124] Alternatively, the clamping unit can also be replaced by a linear circulating guide rail or a multi-row telescopic buffer rack, which are also common alternatives in this field.

[0125] A smart sorting system for clothing goods oriented to combined orders is provided. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of a smart sorting method for clothing goods oriented to combined orders.

Claims

1. A method for intelligent sorting of apparel goods for combined orders, characterized in that, The method includes: Images of clothing items on the conveyor are collected and preprocessed to generate images to be matched. At the same time, the current order and the corresponding user identifier are obtained. The current order contains multiple attributes of each clothing item. Retrieve the user's historical order data based on the user identifier, input the historical order data into the pre-trained prediction model to obtain the predicted order and the corresponding confidence score; compare the current order with the predicted order to generate the attention score of each attribute. Based on the attention of each attribute, at least one key attribute is determined, and the corresponding feature descriptor is read from the pre-built feature library based on the key attribute to generate a key template. The matching threshold of the key template is adjusted according to the confidence level; a basic template is generated for each product in the current order. Based on the type of key attributes, key templates are used and matching filters are performed in the image to be matched according to the matching threshold to obtain candidate regions and matching scores for each candidate region. Based on the basic templates of each product, each candidate region is finely matched, the matching score of each candidate region is updated, and the spatial relationship between each candidate region is combined to determine the association result between each candidate region and each product in the current order; based on the association result, the corresponding clamping unit is controlled to perform gripping and packaging actions.

2. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 1, characterized in that, The process of generating the image to be matched includes: Obtain a background reference image of the conveyor mechanism under no-load conditions; Background subtraction technology is used to perform pixel-by-pixel difference processing between the acquired image and the background reference image to generate a difference image; The difference image is segmented using a preset separation threshold to extract candidate product regions from the acquired image; The candidate regions for the products are deblurred to obtain the image to be matched.

3. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 1, characterized in that, The prediction model is a sequence prediction model pre-trained using historical order data from multiple users. The process of obtaining the predicted order and its corresponding confidence level includes: Based on the user identifier, retrieve the user's historical order data, which includes the attributes of each clothing item purchased by the user in the past; among which, the attributes include at least category, color, size and pattern; Historical order data is input into a pre-trained prediction model. The prediction model outputs predicted orders and their corresponding confidence scores. Each predicted order contains at least one predicted product, and each predicted product has corresponding attributes.

4. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 3, characterized in that, The process for determining the attention level of each attribute includes: The system pairs each item in the current order with each predicted item in the predicted order to obtain paired items. Each paired item is then compared for differences in each attribute to generate an attribute difference score. Specifically, for category, color, and pattern type attributes, if the attribute value of the item in the current order is the same as its attribute value in the predicted order, the attribute difference score is zero; otherwise, it is set to one. For size-ordered numerical attributes, sizes are mapped to numerical values ​​in ascending order to obtain size values. The difference between the largest and smallest size values ​​is calculated as the maximum size difference. The absolute value of the difference between the size value of the item in the current order and the size value in the predicted order is calculated and then divided by the maximum size difference score to obtain the attribute difference score for that size attribute. Assign a preset base weight to each attribute; From the user's historical order data, we statistically analyze the value distribution of each attribute of the clothing items purchased by the user in the past; for category, color, and pattern type attributes, we use information entropy to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute; for size ordered numerical attributes, we use variance to quantify the concentration of the attribute values ​​to obtain the purchase stability of the attribute. For each attribute, calculate the difference between the value one and the purchase stability of that attribute to obtain the instability of that attribute; multiply the preset modulation strength coefficient and the instability as the modulation amount; subtract the modulation amount from the value one to obtain the modulation coefficient; multiply the preset base weight of that attribute by the corresponding modulation coefficient to obtain the attribute weight of that attribute. Calculate the product of the attribute weight and the corresponding attribute difference degree for each attribute, and use it as the weighted difference degree of each product on each attribute; sum the weighted difference degrees of the same attribute on each product to obtain the total attribute difference value of each attribute; sum and normalize the total attribute difference value of each attribute to obtain the attention degree of each attribute.

5. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 1, characterized in that, The pre-built feature library contains feature descriptors for each clothing item. Each clothing item corresponds to a set of feature descriptors, which include at least the color feature descriptor, texture feature descriptor, and depth feature descriptor for that clothing item. The color feature descriptor indicates the hue range, saturation range, and lightness range of the clothing item in the HSV color space. The texture feature descriptor indicates the local texture pattern distribution histogram of the clothing item. The depth feature descriptor indicates the depth feature vector of the clothing item. The process of determining the matching threshold of the key template includes: Obtain the preset baseline threshold and preset relaxation coefficient; Based on the level of attention of each attribute, select at least one attribute with the highest level of attention as the key attribute; Based on the key attributes, the feature descriptors of the corresponding products are read from the pre-built feature library to generate key templates; if the key attributes contain two attributes, then two corresponding key templates are generated respectively. The difference between the numerical value and the confidence level is used as the unreliability level; the product of the preset relaxation coefficient and the unreliability level is used as the relaxation amount. The difference between the first value and the relaxation amount is calculated to obtain the adjustment factor; the product of the preset benchmark threshold and the adjustment factor is calculated as the adjusted matching threshold.

6. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 5, characterized in that, The process of obtaining the candidate regions and their matching scores includes: The image to be matched is divided into multiple matching units, each of which is a pixel or a sliding window region; If the key attribute is color, then the matching unit is pixel, and the corresponding key template is recorded as the color template; extract the color feature descriptor containing the target color in the color template; perform color matching between each pixel in the image to be matched and the color template, calculate the membership degree of each pixel to the target color, and use it as the matching response result of each pixel; If the key attribute is a pattern, the matching unit is a sliding window region, and the corresponding key template is recorded as a texture template. The image to be matched is traversed in a sliding window manner, and the texture feature descriptors of each window region are extracted in turn. The texture feature matching technology is used to calculate the chi-square distance between the texture feature descriptors of each window region and the texture template, and the chi-square distance is converted into similarity as the matching response result of each window region. The matching response results are filtered according to the matching threshold. The positions that meet the preset conditions after filtering are divided into candidate regions. The matching score of each candidate region is determined by combining the matching response results.

7. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 6, characterized in that, The step of filtering the matching response results according to the matching threshold, dividing the filtered positions that meet the preset conditions into candidate regions, and determining the matching score of each candidate region based on the matching response results includes: Matching units whose matching response results reach the matching threshold are marked as valid units; adjacent valid units are connected and aggregated to form connected regions; the circumscribed regions of each connected region are determined as candidate regions; for each candidate region, the arithmetic mean of the matching response results of each matching unit contained in the candidate region on the same key attribute is calculated as the matching component of the candidate region on the same key attribute. If a candidate region has both color matching results and texture matching results, the matching components of all key attributes contained in the candidate region are weighted and fused according to the preset first weight to generate the matching score of the candidate region.

8. The intelligent sorting method for apparel goods oriented towards combined orders according to claim 5, characterized in that, The basic template is a corresponding deep feature descriptor read from a pre-built feature library based on the product identifiers carried by each product in the current order; The process of performing fine-grained matching on each candidate region based on the basic template of each product and updating the matching score of each candidate region includes: Extract candidate image patches corresponding to each candidate region from the image to be matched; Each candidate image patch is finely matched with the base template of each product in the current order, and the depth feature descriptor of each candidate image patch is extracted. The cosine distance between the depth feature descriptor of each candidate image patch and the depth feature descriptor of each base template is calculated, and the cosine distance is mapped to a value between zero and one, which is used as the fine matching score between each candidate image patch and each base template; where the cosine distance and the fine matching score are negatively correlated. For each candidate region, the highest score is selected from the fine matching scores of the candidate region and each product, and this score is used as the updated matching score for the candidate region. The product corresponding to the highest score and its product identifier are recorded. If the highest score is lower than the preset fine matching threshold, the candidate region is determined not to belong to any product in the current order and is removed.

9. A method for intelligent sorting of apparel goods for combined orders according to claim 8, characterized in that, The process for determining the association result includes: Determine the center point of each candidate region; calculate the pairwise distance between the center points of each candidate region; calculate the arithmetic mean of all pairwise distances as the average spacing between candidate regions. The ratio of the average spacing to the preset reference spacing is calculated as the spacing ratio; the spacing ratio is mapped to a value between zero and one using an exponential decay method to obtain the location clustering degree. Calculate the arithmetic mean of the updated matching scores for all candidate regions as the average refinement score; The average fine score and the location clustering degree are each assigned a preset second weight, and then weighted and fused to obtain the comprehensive confidence score. If the overall confidence level reaches or exceeds the preset decision threshold, each candidate region is determined to be successfully associated with the current order, and the association result corresponding to each candidate region is output. The association result is used to indicate whether each candidate region belongs to the current order and which product in the current order it belongs to. If the overall confidence level is lower than the preset decision threshold, the association is deemed to have failed, and the process is transferred to the exception handling procedure.

10. A smart sorting system for apparel goods oriented towards combined orders, characterized in that, The system includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 9.

Images