Intelligent warehouse management method and system based on visual feature parameter detection

Abstract

The invention discloses an intelligent warehouse management method and system based on visual feature parameter detection, and belongs to the field of intelligent warehouse management. The method comprises the following steps: data acquisition: acquiring white spirit labels, bottle body contours and goods allocation identification images through image acquisition equipment, acquiring goods allocation coordinates through a 3D sensor, and synchronously recording time; data processing: preprocessing the images, constructing a sample library containing white spirit categories, label states and the like, and training a CNN model; analyzing data, inputting image features, and calculating five types of core parameters in combination with coordinates; processing results, updating goods allocation and inventory information when the parameters reach the standard, and triggering alarm and generating suggestions when the parameters do not reach the standard. The system comprises data acquisition, processing, analysis and result processing modules. The method improves the white spirit feature recognition precision, achieves the precise management and control of the goods location and the inventory, timely responds to the abnormality, reduces the manual intervention, and improves the storage intelligent level and efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent warehouse management, and in particular to an intelligent warehouse management method and system based on visual feature parameter detection. Background Technology

[0002] With the rapid development of the liquor industry and the increasing diversification of liquor categories, different brands, alcohol content, and bottle types of liquor place higher demands on the precision and efficiency of warehouse management. Currently, the liquor warehouse management field is gradually shifting from traditional manual inventory to semi-automation and intelligent systems. The industry is widely introducing hardware facilities such as image acquisition equipment and sensors to assist in recording cargo information. At the same time, the application of deep learning technology in image recognition and feature extraction is becoming increasingly widespread, providing a technological foundation for cargo identification and management through visual features. The overall level of intelligence in warehouse management is gradually improving.

[0003] However, traditional liquor warehousing management typically relies on manual methods for goods identification, location verification, and inventory counting. This is not only inefficient but also prone to problems such as misclassification, missing or damaged labels, incorrect bottle types, or misplaced storage locations due to subjective factors. With the development of visual technology, some systems have begun to use image recognition methods for goods inspection, but these still suffer from limitations such as single feature dimensions, poor environmental adaptability, and insufficient recognition accuracy. In particular, they lack the ability to comprehensively judge and dynamically analyze multiple feature parameters, failing to effectively support the needs of refined, real-time warehouse management and struggling to achieve highly reliable intelligent warehouse status perception and decision support. Summary of the Invention

[0004] The purpose of this invention is to overcome one or more shortcomings of the prior art and provide an intelligent warehouse management method and system based on visual feature parameter detection.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] In a first aspect, this application provides an intelligent warehouse management method based on visual feature parameter detection, comprising the following steps:

[0007] S1. Data Acquisition: Acquire images of the liquor label, bottle outline, and location markings using image acquisition equipment; acquire spatial coordinate data of the location using 3D sensors; and simultaneously record acquisition time information.

[0008] S2. Data Processing: Preprocess the images acquired in S1, construct a sample library containing liquor categories, label status, bottle type features, and storage location, and train a CNN model for feature recognition;

[0009] S3. Data Analysis: Input the preprocessed image into the CNN model to extract the characteristics of liquor category, label integrity, and bottle type matching. Combine the location coordinate data to calculate five feature parameters: liquor category recognition accuracy, label integrity recognition rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate.

[0010] S4. Result Processing: Based on the feature parameters obtained in S3, if the category identification accuracy, label integrity identification rate, and bottle type matching degree all reach the preset threshold and the location coordinate deviation is within the allowable range, then update the location binding information and inventory data; if any parameter fails to meet the standard, trigger an abnormal alarm and generate processing suggestions.

[0011] In some embodiments, step S2, which involves constructing the sample library and training the CNN model, includes the following sub-steps:

[0012] S2.1 Sample Library Construction: Collect images of baijiu samples of different brands, alcohol content, and vintages, and label them with category labels, label integrity labels, bottle type labels, and location coordinate labels;

[0013] S2.2 Model Training: Using transfer learning, based on a pre-trained CNN model, fine-tuning is performed using sample database data. The model's ability to extract features of category, label, and bottle type is optimized through a multi-task loss function until the model's overall recognition accuracy on the validation set reaches the target.

[0014] In some embodiments, the multi-task loss function in step S2.2 includes category identification loss, label integrity identification loss, and bottle type matching loss. The influence weight of each loss item is adjusted by dynamic weight coefficients so that the model prioritizes optimizing the identification accuracy of key features of warehouse management.

[0015] In some embodiments, in step S3, when calculating the accuracy rate of liquor category identification, the liquor category identification result, the management weight corresponding to the category, and the collection time of each identification are obtained. The category weight and time decay coefficient are weighted and statistically analyzed by a preset accuracy algorithm model to obtain the liquor category identification accuracy rate.

[0016] In some embodiments, in step S3, when calculating the cargo location coordinate deviation value, the actual three-dimensional coordinates of the cargo placement, the preset three-dimensional coordinates of the cargo location, and the deviation weight coefficients of each coordinate axis are obtained by acquiring the three-dimensional coordinates of the cargo placement collected by the 3D sensor, and the spatial difference comprehensive calculation is performed on the deviation of each axis coordinate according to the preset deviation value algorithm model to obtain the cargo location coordinate deviation value.

[0017] In some embodiments, in step S3, when calculating the inventory quantity statistical error rate, the inventory quantity is calculated by obtaining the model-identified quantity of inventory counts within the statistical period, the actual inventory quantity manually verified, and the time decay coefficient. The ratio is calculated by weighting the quantity difference of each count with the weighted sum of the actual inventory quantity using a preset error rate algorithm model, and then obtaining the inventory quantity statistical error rate.

[0018] In some embodiments, the triggering conditions for the abnormal alarm in step S4 include at least one of the following: the accuracy rate of liquor category identification is lower than a preset threshold, the label integrity identification rate is lower than a preset threshold, the bottle type matching degree does not reach a preset threshold, the location coordinate deviation value exceeds the allowable range, and the inventory quantity statistical error rate exceeds a preset threshold.

[0019] Secondly, this application provides an intelligent warehouse management system based on feature parameter detection, comprising:

[0020] The data acquisition module is used to acquire images of the liquor, the coordinates of the storage location, and time information through image acquisition equipment and 3D sensors;

[0021] The data processing module is used to preprocess the acquired images, build a sample library, and train a CNN model.

[0022] The data analysis module is used to extract the characteristics of baijiu through a CNN model and calculate the accuracy of baijiu category identification, label integrity identification rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate.

[0023] The results processing module is used to update the location and inventory information based on the data analysis results, trigger anomaly alarms, and generate handling suggestions.

[0024] In some embodiments, the data acquisition module includes at least:

[0025] High-definition camera used to capture images of liquor labels and bottles;

[0026] 3D vision sensors are used to acquire the three-dimensional coordinates of the cargo location;

[0027] The time synchronization unit is used to ensure that the timestamps of the image and coordinate data are consistent.

[0028] In some embodiments, the result processing module includes at least:

[0029] The information update unit is used to update the location and inventory data when the parameters meet the standards.

[0030] Anomaly detection unit is used to compare parameters with preset thresholds;

[0031] The alarm unit is used to trigger an audible and visual alarm and push information to the management terminal when parameters do not meet the standards.

[0032] The beneficial effects of this invention are:

[0033] (1) By combining visual acquisition and 3D sensing with image preprocessing and CNN model training technology, the recognition accuracy of liquor category, label status and bottle type is improved, and the impact of environmental interference on warehouse feature recognition is effectively reduced.

[0034] (2) By calculating multiple characteristic parameters and combining them with an anomaly alarm mechanism, the precise control of storage location matching and inventory can be achieved. It can promptly locate the problem of parameters not meeting the standards and push processing suggestions to ensure that the storage status is controllable.

[0035] (3) By coordinating data collection, processing, analysis and result processing modules, a closed loop of warehouse management is constructed to reduce manual intervention and improve the level of intelligence and overall efficiency of the entire process of liquor warehouse management. Attached Figure Description

[0036] Figure 1 A flowchart illustrating the specific steps of the intelligent warehouse management method provided in this embodiment of the invention;

[0037] Figure 2 This invention provides an operational diagram illustrating the entire process of intelligent warehouse management, from data acquisition and processing to analysis and result processing.

[0038] Figure 3 This is a diagram of the dual-branch feature extraction and attention-weighted fusion structure of the CNN model for liquor storage provided in this embodiment of the invention.

[0039] Figure 4 This is a training diagram of the entire process of "sample input - pre-training - fine-tuning - loss calculation - parameter update" for the CNN model for liquor storage provided in this embodiment of the invention.

[0040] Figure 5 A block diagram of the intelligent warehouse management system provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0042] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0043] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0044] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0045] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0046] In the traditional warehousing management of the liquor industry, manual methods are generally relied upon for goods identification, location verification, and inventory counting. This model is not only inefficient and costly, but also highly susceptible to problems caused by subjective factors such as staff fatigue and experience differences. These problems include misidentification of liquor categories, missing or damaged labels, incorrect bottle types, or misplaced storage locations, directly impacting the accuracy of inventory data and the efficiency of inbound and outbound operations. With the development of computer vision technology, some modern warehouses have begun to introduce image recognition-based detection systems in an attempt to automate goods management. However, these systems often suffer from limitations such as single feature dimensions, poor environmental adaptability, and insufficient recognition accuracy. In particular, they lack the ability to comprehensively judge and dynamically analyze multiple feature parameters (such as visual appearance and spatial location), failing to effectively support the needs of refined and real-time warehousing management. Furthermore, existing methods are still imperfect in key technical aspects such as storage location spatial information fusion, time-decay weighted statistics, and multi-task model optimization, making it difficult to achieve highly reliable intelligent state perception and decision support. In view of this, this application provides an intelligent warehouse management method and system based on visual feature parameter detection, so as to realize a warehouse management method that integrates multi-source visual features and three-dimensional coordinate data, and has the ability to dynamically evaluate, intelligently alarm and generate optimization suggestions, thereby comprehensively improving the automation level, accuracy and operational efficiency of liquor warehouse management.

[0047] Example 1

[0048] Reference Figure 1 The first aspect of this application provides an intelligent warehouse management method based on visual feature parameter detection, the method comprising the following steps:

[0049] S1. Data Acquisition: Acquire images of the liquor label, bottle outline, and location markings using image acquisition equipment; acquire spatial coordinate data of the location using 3D sensors; and simultaneously record acquisition time information.

[0050] S2. Data Processing: Preprocess the images acquired in S1 (deblurring, noise reduction, illumination correction), construct a sample library containing liquor categories, label status, bottle type features, and storage location, and train a CNN model for feature recognition;

[0051] S3. Data Analysis: Input the preprocessed image into the CNN model to extract the characteristics of liquor category, label integrity, and bottle type matching. Combine the location coordinate data to calculate five feature parameters: liquor category recognition accuracy, label integrity recognition rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate.

[0052] S4. Result Processing: Based on the feature parameters obtained in S3, if the category identification accuracy, label integrity identification rate, and bottle type matching degree all reach the preset threshold and the location coordinate deviation is within the allowable range, then update the location binding information and inventory data; if any parameter fails to meet the standard, trigger an abnormal alarm and generate processing suggestions.

[0053] Step S2, which involves constructing the sample library and training the CNN model, includes the following sub-steps:

[0054] S2.1 Sample Library Construction: Collect images of baijiu samples of different brands, alcohol content, and years, and label them with category labels (such as brand, alcohol content), label integrity labels (such as intact, worn, detached), bottle type labels (such as cylindrical bottle, square bottle), and location coordinate labels.

[0055] S2.2 Model Training: Using transfer learning, based on a pre-trained CNN model, fine-tuning is performed using sample database data. The model's ability to extract features of category, label, and bottle type is optimized through a multi-task loss function until the model's overall recognition accuracy on the validation set reaches the target.

[0056] In step S3, the accuracy rate of liquor category identification is calculated by obtaining the liquor category identification result (correct or incorrect), the management weight corresponding to the category (set according to the importance of goods), and the collection time of each identification. The category weight and the time decay coefficient (to reduce the interference of historical data on the current result) are weighted and statistically analyzed to obtain the liquor category identification accuracy rate. This accuracy rate can reflect the accuracy of the model in identifying liquor categories of different importance.

[0057] The specific formula for calculating the accuracy of liquor category identification is as follows:

[0058] ;

[0059] in, denoted as , where n is the total number of baijiu (Chinese liquor) categories identified. The result of identifying the i-th type of liquor (correct or incorrect). Let be the category weight of the i-th type of baijiu. Let be the collection time of the i-th liquor, T be the current time, and α be the time decay coefficient.

[0060] In step S3, the deviation value of the cargo location coordinates is calculated by obtaining the actual three-dimensional coordinates of the cargo placement, the preset three-dimensional coordinates of the cargo location, and the deviation weight coefficients of each coordinate axis (x, y, z axes) collected by the 3D sensor. The deviation of each axis coordinate is assigned an appropriate weight and a comprehensive calculation of spatial differences is performed to obtain the deviation value of the cargo location coordinates. This deviation value can quantify the spatial distance difference between the actual position of the cargo and the preset cargo location.

[0061] The specific formula for calculating the deviation value of the cargo location coordinates is as follows:

[0062] ;

[0063] in, This represents the deviation value of the cargo location coordinates. The deviation between the actual x-axis coordinates and the target x-axis coordinates y represents the deviation between the actual y-axis coordinate and the target coordinate, kx, ky, and kz are the deviation weighting coefficients for the x, y, and z axes, respectively, and kxy is the coefficient of the xy-axis deviation interaction term.

[0064] In step S3, the inventory quantity statistical error rate is calculated by obtaining the inventory quantity identified by the model from multiple inventory counts within the statistical period, the actual inventory quantity verified by manual review, and the time decay coefficient (the weight of recent inventory count data is higher than that of distant data). The inventory quantity statistical error rate is obtained by weighting the quantity difference of each inventory count and calculating the ratio with the weighted sum of the actual inventory quantity. This error rate can assess the accuracy level of the current inventory statistics.

[0065] The specific formula for calculating the inventory quantity statistical error rate is as follows:

[0066]

[0067] in, The error rate for inventory quantity statistics is denoted by m, where m is the number of inventory counts within the statistical period. Let be the number of models identified in the t-th inventory count. Let λ be the number of manual checks performed during the t-th inventory count, T be the current inventory count time, and λ be the time decay coefficient.

[0068] The triggering conditions for the abnormal alarm in step S4 include at least one of the following: the accuracy rate of liquor category identification is lower than the preset threshold, the label integrity identification rate is lower than the preset threshold, the bottle type matching degree does not reach the preset threshold, the location coordinate deviation value exceeds the allowable range, and the inventory quantity statistics error rate exceeds the preset threshold.

[0069] In step S2.2, the multi-task loss function includes category identification loss, label integrity identification loss, and bottle type matching loss. The influence weight of each loss item is adjusted by dynamic weight coefficients so that the model prioritizes optimizing the identification accuracy of key features of warehouse management.

[0070] Example 2

[0071] like Figure 2 As shown, three types of key images are acquired directionally through image acquisition equipment: liquor label image (clearly showing brand logo, alcohol content and other details), bottle outline image (completely recording bottle structure and stacking status), and cargo location identification image (accurately acquiring cargo location code information).

[0072] Simultaneously, 3D sensors are used to collect spatial three-dimensional coordinate data of the cargo location, recording the actual spatial position of the goods. A time synchronization unit is configured to add a unified timestamp to each set of image data and coordinate data, ensuring the correlation between "image-coordinate-time". During the collection process, blurry images and invalid coordinates that are outside the cargo location range are filtered out to form a complete original dataset.

[0073] Use a deblurring algorithm to restore the edges of label text and the details of bottle outline, and eliminate image blurring caused by dust obscuring or equipment shaking;

[0074] The algorithm combines Gaussian filtering and median filtering to remove noise from the image (such as isolated noise points formed by shelf stains) while preserving the purity of the feature regions.

[0075] Histogram equalization is used to adjust the grayscale distribution of the image, balance the brightness difference between the bright areas on the top of the shelf and the shadow areas on the bottom, avoid the interference of uneven lighting on subsequent feature recognition, and finally output high-quality clean image data.

[0076] A dedicated sample library was built based on images: the samples cover the differences in the categories of liquor (label design and bottle type differentiation for different brands and alcohol content), label status (intact, slightly worn, heavily worn, detached), bottle type characteristics (cylindrical, square, flat) and storage location (different shelf areas, different shelf heights). Each sample is labeled with the corresponding category information, label integrity level, bottle type and storage location code.

[0077] The samples are enhanced by methods such as rotation (simulating the deviation of the acquisition angle), scaling (simulating the difference in distance between the goods and the equipment), and brightness adjustment (simulating changes in lighting) to expand the sample coverage and avoid overfitting in subsequent model training.

[0078] The CNN model is trained based on a sample library. The pre-trained CNN model is fine-tuned using sample library data. The model's ability to extract features of category, label, and bottle type is optimized through a multi-task loss function until the model's overall recognition accuracy on the validation set reaches the preset requirements.

[0079] The preprocessed images are batch-input into the trained CNN model according to their types. After the label images are processed by the model, the category feature code and label integrity feature value are output. The bottle outline image outputs the bottle shape feature code and stacking state feature value. The location identification image outputs the location code feature vector.

[0080] Based on the location coordinate data collected by 3D sensors, five types of parameters are calculated: liquor category recognition accuracy (the proportion of goods whose category is correctly identified by the statistical model out of the total number of identified goods), label integrity recognition rate (the proportion of goods whose label status is correctly judged by the statistical model), bottle type matching degree (measuring the degree of fit between the bottle type identified by the model and the actual bottle type), location coordinate deviation value (calculating the spatial distance between the actual placement coordinates of the goods and the preset coordinates of the location), and inventory quantity statistics error rate (the proportion of the difference between the inventory quantity counted by the model and the actual inventory quantity verified by manual verification).

[0081] The system executes management actions based on the characteristic parameters: if the category identification accuracy, label integrity identification rate, and bottle type matching degree all reach the preset threshold, and the location coordinate deviation is within the allowable range, the system will automatically update the binding information between the location and the goods, synchronously correct the inventory data in the warehouse management system, generate a report containing the analysis time, the number of locations covered, and the specific values ​​of each parameter, and archive it.

[0082] If any parameter fails to meet the standard (such as the category identification accuracy rate being lower than the threshold or the location coordinate deviation being outside the range), an abnormal alarm will be triggered immediately, alerting on-site management personnel through audible and visual signals. At the same time, a report containing the abnormality type (such as category identification deviation or misplacement of goods), abnormal location, abnormal parameter value, and handling suggestions (such as re-collecting label images for secondary identification or adjusting the goods placement location to calibrate coordinates) will be generated and pushed to the management terminal.

[0083] After the management personnel have processed the report, the system re-collects data on the abnormal storage locations, performs feature extraction and parameter calculation, verifies whether the abnormality has been resolved, and ensures that warehouse management meets the accuracy requirements.

[0084] Example 3

[0085] like Figure 3 As shown, the model takes preprocessed images related to baijiu (Chinese liquor) as input and adopts an architecture of "dual-branch parallel extraction + attention-weighted fusion" to achieve accurate feature capture and efficient utilization.

[0086] One branch focuses on extracting detailed features from liquor label images. By using depthwise separable convolutional layers to reduce computation while preserving key details such as category identification and alcohol content on the label, the attention module redistributes the feature channel weights to enhance features in areas such as the brand logo and text edges. Finally, it outputs a label detail feature vector with uniform dimensions. This vector directly supports subsequent category identification and label integrity judgment, and forms a corresponding association with the category labels and label integrity labels (such as complete, worn, and detached) marked when building the sample library.

[0087] Another branch performs global feature extraction on the bottle outline and packaging image. It solves the gradient vanishing problem in deep networks by using multiple sets of residual convolutional blocks, effectively capturing global information such as bottle structure (e.g., the outline differences between cylindrical and square bottles) and packaging stacking status. The two-dimensional feature map is transformed into a one-dimensional global feature vector of bottle / packaging through a global average pooling layer. The dimension design of this vector is adapted to the feature differentiation requirements of different bottle types, and it matches with the bottle type labels marked in the sample library, providing a feature basis for subsequent bottle type matching degree calculation.

[0088] After the two feature vectors are fed into the feature fusion layer, they are integrated using an attention-weighted fusion module.

[0089] During the fusion process, based on the priority requirements of warehouse management for different features (such as category identification and label status judgment being more critical for accurate warehouse management), higher dynamic weights are assigned to the label detail feature vectors, and adaptive weights are assigned to the bottle type / packaging global feature vectors. The fused feature vector is obtained through weighted calculation. This vector covers both detailed and global features, avoiding recognition bias caused by a single feature dimension.

[0090] Dynamic weight adjustment (the multi-task loss function optimizes the recognition accuracy of key warehouse features by prioritizing dynamic weights) ensures that features play a dominant role in the final result.

[0091] After the fused feature vectors are fed into the output layer, two types of results are output: one is the category / label / bottle type classification result, which outputs the category recognition probability, label integrity status probability, and bottle type probability through a fully connected layer and a softmax function. This result can then be combined with the formula:

[0092] Calculate the accuracy rate of liquor category identification, among which, The value should be chosen based on the correctness of the category identification in the classification results. Based on the importance of category management;

[0093] Another type is the feature coordinate positioning result, which outputs the specific coordinate information of the liquor in the storage location through the regression layer, providing a positional basis for the subsequent calculation of the storage location coordinate deviation value (combining the x, y, z axis deviation and weight coefficient calculation), realizing the collaborative output of "feature recognition-coordinate positioning", and supporting the dual needs of accurate category identification and dynamic storage location matching in warehouse management.

[0094] During the model training phase, a transfer learning approach is adopted based on this architecture. The pre-trained CNN model is fine-tuned using sample library data. The parameters of each branch and fusion layer are optimized through a multi-task loss function (including category recognition loss, label integrity recognition loss, and bottle type matching loss) until the comprehensive accuracy of category recognition, label judgment, and bottle type matching on the validation set reaches the standard. This ensures that the model can stably output high-quality features, classification, and localization results, providing technical support for feature parameter calculation and management decisions in intelligent baijiu warehousing.

[0095] Example 4

[0096] like Figure 4 As shown, the model training is based on the sample library as input. The sample library needs to cover dimensions such as liquor category, label status, bottle type features and storage location. Each sample needs to be labeled with storage location coordinates (such as spatial coordinate information corresponding to different shelf layers and columns).

[0097] After the sample is input, the model first enters the pre-training stage. The model weights are initialized based on the ImageNet dataset, and an appropriate learning rate is adopted to enable the model to initially have the ability to extract general image features. In this stage, there is no need to adjust the parameters of the underlying convolutional layers. The model only learns basic features such as edges and textures through general data, laying the foundation for subsequent scene-specific fine-tuning.

[0098] After pre-training is completed, the fine-tuning stage begins. At this stage, the bottom convolutional layers of the model need to be frozen (to retain the ability to extract general features), and only the parameters of the top fully connected layers are updated.

[0099] During fine-tuning, the learning rate was adjusted to a value suitable for the warehousing scenario. Simultaneously, samples were input in batches according to the "category-location-environment" combination, enabling the model to specifically learn the characteristic performance of different categories of baijiu in different locations and environments. This stage employs a multi-task loss function:

[0100] in, As a dynamic weighting coefficient, it can be adjusted according to warehouse management needs, such as when the accurate location of the storage location is given higher priority, by appropriately increasing the weight of loss items related to bottle type and storage location, so as to ensure that the model prioritizes the extraction accuracy of key features such as storage location matching and bottle type recognition.

[0101] The loss calculation process requires real-time monitoring of three types of losses: category identification loss, which measures the model's prediction bias for the liquor category;

[0102] To mitigate the loss in label integrity identification, samples with worn labels are given higher weights.

[0103] Bottle shape matching loss is optimized by comparing positive and negative pressure samples to improve the distinguishability of bottle shape features.

[0104] The model parameters are updated using the backpropagation algorithm based on the loss value, gradually reducing the loss difference between the training set and the validation set, thus avoiding model overfitting.

[0105] Once the overall accuracy of category identification, label judgment, and bottle type matching on the validation set meets the standard, the model training is completed and the final model file is output. This model can be directly used in the warehouse data analysis process.

[0106] The trained model needs to support feature parameter calculation, such as the calculation of the cargo location coordinate deviation: combining the actual cargo location coordinates collected by the 3D sensor with the feature coordinate positioning results output by the model, and substituting them into the formula:

[0107]

[0108] in, These represent the deviations between the actual coordinates and the target coordinates of the x, y, and z axes, respectively. kx, ky, and kz are the weighting coefficients for each axis deviation, and kxy is the coefficient of the xy axis deviation interaction term. This formula accurately quantifies the spatial difference between the placement location of the goods and the preset storage location.

[0109] The model can also assist in calculating the inventory quantity statistical error rate. Multiple inventory checks are conducted within the statistical period, and the inventory quantity identified by the model and the manually verified quantity are substituted into the formula:

[0110]

[0111] By combining the time decay coefficient λ, the impact of recent inventory data is highlighted, providing a quantitative basis for assessing the accuracy of inventory management.

[0112] After the model is deployed, it needs to work in conjunction with the data analysis module and result processing module of the warehousing system.

[0113] The data analysis module calls the model to extract features and calculate parameters. The result processing module then triggers information updates or abnormal alarms based on whether the parameters meet the standards (such as whether the deviation value of the storage location coordinates is within the allowable range or whether the inventory error rate is lower than the threshold). For example, if the parameters do not meet the standards, the management personnel are alerted through an audible and visual alarm device, and at the same time, a processing suggestion containing the abnormal storage location and parameter deviation value is pushed, forming a collaborative link of "model training - parameter calculation - management decision".

[0114] Example 5

[0115] Reference Figure 5 The second aspect of this application provides an intelligent warehouse management system based on feature parameter detection to implement the above method. The system includes:

[0116] The data acquisition module is used to acquire images of the liquor, the coordinates of the storage location, and time information through image acquisition equipment and 3D sensors;

[0117] The data processing module is used to preprocess the acquired images, build a sample library, and train a CNN model.

[0118] The data analysis module is used to extract the characteristics of baijiu through a CNN model and calculate the accuracy of baijiu category identification, label integrity identification rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate.

[0119] The results processing module is used to update the location and inventory information based on the data analysis results, trigger anomaly alarms, and generate handling suggestions.

[0120] In some embodiments, the data acquisition module includes a high-definition camera, a 3D vision sensor, and a time synchronization unit. The high-definition camera is used to acquire images of the liquor label and bottle, the 3D vision sensor is used to acquire the three-dimensional coordinates of the storage location, and the time synchronization unit is used to ensure that the timestamps of the images and coordinate data are consistent.

[0121] In some embodiments, the result processing module includes an information update unit, an anomaly judgment unit, and an alarm unit. The information update unit is used to update the location and inventory data when the parameters meet the standards. The anomaly judgment unit is used to compare the parameters with a preset threshold. The alarm unit is used to trigger an audible and visual alarm and push information to the management terminal when the parameters do not meet the standards.

[0122] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0123] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An intelligent warehouse management method based on visual feature parameter detection, characterized in that, Includes the following steps: S1. Data Acquisition: Acquire images of the liquor label, bottle outline, and location markings using image acquisition equipment; acquire spatial coordinate data of the location using 3D sensors; and simultaneously record acquisition time information. S2. Data Processing: Preprocess the images acquired in S1, construct a sample library containing liquor categories, label status, bottle type features, and storage location, and train a CNN model for feature recognition; S3. Data Analysis: Input the preprocessed image into the CNN model to extract the characteristics of liquor category, label integrity, and bottle type matching. Combine the location coordinate data to calculate five feature parameters: liquor category recognition accuracy, label integrity recognition rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate. S4. Result Processing: Based on the feature parameters obtained in S3, if the category identification accuracy, label integrity identification rate, and bottle type matching degree all reach the preset threshold and the location coordinate deviation is within the allowable range, then update the location binding information and inventory data; if any parameter fails to meet the standard, trigger an abnormal alarm and generate processing suggestions.

2. The method according to claim 1, characterized in that, Step S2, which involves building the sample library and training the CNN model, includes the following sub-steps: S2.1 Sample Library Construction: Collect images of baijiu samples of different brands, alcohol content, and vintages, and label them with category labels, label integrity labels, bottle type labels, and location coordinate labels; S2.2 Model Training: Using transfer learning, based on a pre-trained CNN model, fine-tuning is performed using sample database data. The model's ability to extract features of category, label, and bottle type is optimized through a multi-task loss function until the model's overall recognition accuracy on the validation set reaches the target.

3. The method according to claim 2, characterized in that, In step S2.2, the multi-task loss function includes category identification loss, label integrity identification loss, and bottle type matching loss. The influence weight of each loss item is adjusted by dynamic weight coefficients so that the model prioritizes optimizing the identification accuracy of key features of warehouse management.

4. The method according to claim 1, characterized in that, In step S3, when calculating the accuracy of liquor category identification, the liquor category identification result, the management weight corresponding to the category, and the collection time of each identification are obtained. The category weight and time decay coefficient are weighted and statistically analyzed by a preset accuracy algorithm model to obtain the liquor category identification accuracy.

5. The method according to claim 1, characterized in that, In step S3, when calculating the cargo location coordinate deviation value, the actual three-dimensional coordinates of the cargo placement, the preset three-dimensional coordinates of the cargo location, and the deviation weight coefficients of each coordinate axis are obtained by acquiring the three-dimensional coordinates of the cargo placement collected by the 3D sensor, and the spatial difference comprehensive calculation is performed according to the preset deviation value algorithm model to assign appropriate weights to the coordinate deviations of each axis and obtain the cargo location coordinate deviation value.

6. The method according to claim 1, characterized in that, In step S3, when calculating the inventory quantity statistical error rate, the model-identified quantity of inventory counts, the actual inventory quantity verified manually, and the time decay coefficient are obtained within the statistical period. The weighted sum of the quantity difference of each count and the actual inventory quantity is calculated using a preset error rate algorithm model to obtain the inventory quantity statistical error rate.

7. The method according to claim 1, characterized in that, The triggering conditions for the abnormal alarm in step S4 include at least one of the following: the accuracy rate of liquor category identification is lower than the preset threshold, the label integrity identification rate is lower than the preset threshold, the bottle type matching degree does not reach the preset threshold, the location coordinate deviation value exceeds the allowable range, and the inventory quantity statistics error rate exceeds the preset threshold.

8. An intelligent warehouse management system based on feature parameter detection, characterized in that, include: The data acquisition module is used to acquire images of the liquor, the coordinates of the storage location, and time information through image acquisition equipment and 3D sensors; The data processing module is used to preprocess the acquired images, build a sample library, and train a CNN model. The data analysis module is used to extract the characteristics of baijiu through a CNN model and calculate the accuracy of baijiu category identification, label integrity identification rate, bottle type matching degree, location coordinate deviation value, and inventory quantity statistics error rate. The results processing module is used to update the location and inventory information based on the data analysis results, trigger anomaly alarms, and generate handling suggestions.

9. The system according to claim 8, characterized in that, The data acquisition module includes at least: High-definition camera used to capture images of liquor labels and bottles; 3D vision sensors are used to acquire the three-dimensional coordinates of the cargo location; The time synchronization unit is used to ensure that the timestamps of the image and coordinate data are consistent.

10. The system according to claim 8, characterized in that, The result processing module includes at least: The information update unit is used to update the location and inventory data when the parameters meet the standards. Anomaly detection unit is used to compare parameters with preset thresholds; The alarm unit is used to trigger an audible and visual alarm and push information to the management terminal when parameters do not meet the standards.

Images