A precise identification method and system based on intelligent traceability scales

Abstract

This invention discloses a method and system for accurate identification based on an intelligent traceability scale, belonging to the field of data processing and image recognition technology. The method includes: obtaining the unique device identification code of the intelligent traceability scale, binding it with the agricultural market signage and stall signage to complete registration and activation; controlling the intelligent traceability scale to enter learning mode at a designated stall, collecting image data of various commodities; associating the image data with commodity name labeling information input through an interactive interface to construct a local training dataset; using the local training dataset to train a lightweight image recognition model locally on the intelligent traceability scale, obtaining a dedicated image recognition model; switching the intelligent traceability scale to working mode, using the dedicated image recognition model to perform real-time identification of the commodities to be identified on the scale pan, and outputting the commodity identification result.

Description

Technical Field

[0001] This invention relates to the field of data processing and image recognition technology, specifically to a method and system for accurate identification based on an intelligent traceability scale. Background Technology

[0002] In existing applications of smart traceability scales in farmers' markets, product identification typically relies on a general image recognition model deployed on a remote cloud server. Merchants need to photograph products using the smart traceability scale's camera, upload the image data to the cloud, and have the cloud model process it and return the recognition result. This method requires the smart traceability scale to maintain a continuous and stable network connection. Furthermore, the general image recognition model is trained on a large amount of product image data from different scenarios and varieties, making it difficult to accurately adapt to the differences in appearance characteristics of specific products within a single stall. For example, the recognition accuracy of the same type of vegetable fluctuates significantly depending on its growth stage and placement. Existing technical solutions also suffer from the problem of separation between equipment management and deployment processes. The registration and activation of the smart traceability scale is not closely linked to the subsequent activation of the recognition function, resulting in unclear equipment ownership information and the inability of the recognition model to be specifically optimized for the sales characteristics of products at a particular stall. Summary of the Invention

[0003] This invention aims to provide a precise identification method and system based on intelligent traceability scales, in order to solve the problems of existing intelligent traceability scale commodity identification methods relying on remote cloud models, which cannot work in environments without or with weak networks, and the inability of general image recognition models to adapt to the specific appearance features of commodities in a single stall, resulting in low identification accuracy. At the same time, it overcomes the shortcomings of the independent nature of device registration and activation and recognition model training, which makes it difficult to achieve precise association between the device and the stall.

[0004] The objective of this invention can be achieved through the following technical solutions: This invention provides a method and system for accurate identification based on intelligent traceability scales, aiming to solve the problems of low accuracy in commodity identification, insufficient model generalization ability, and high deployment costs of existing intelligent traceability scales in agricultural market scenarios. This invention achieves accurate registration and activation of the device by binding the unique device identification code of the intelligent traceability scale with the agricultural market logo and stall logo, ensuring that each stall's scale uniquely corresponds to the business scenario and avoiding confusion in the identification model. Furthermore, after the intelligent traceability scale is deployed to a designated stall, it is controlled to enter a learning mode. In this mode, image data of various commodities placed on the weighing pan are collected, enabling the collected images to accurately reflect the visual characteristics of the actual commodities sold at the stall (such as lighting, placement angle, and commodity freshness), thereby providing highly relevant local data for subsequent model training.

[0005] Based on the product name labeling information entered through the interactive interface, image data is associated with the product name labeling information to construct a local training dataset. Preferably, the interactive interface can display a candidate list of product names dynamically sorted according to the historical transaction frequency of the farmers' market, reducing the burden of manual input for merchants and improving labeling efficiency and accuracy. Using this local training dataset, a lightweight image recognition model is trained locally on the smart traceability scale to obtain a dedicated image recognition model for that specific stall. Since the lightweight model completes training and inference on the local processor, it does not rely on cloud computing power, significantly reducing network transmission latency and bandwidth costs, while protecting merchant data privacy. The model training process uses a cross-entropy loss function combined with the Adam optimization algorithm; after training convergence, the model parameters are fixed to ensure stable recognition.

[0006] After switching the intelligent traceability scale from learning mode to working mode, the scale uses a dedicated image recognition model to identify the products on the weighing pan in real time and outputs the product identification results. Preferably, during the identification process, if the difference between the highest probability and the second highest probability output by the model is lower than a preset threshold, a recognition ambiguity signal is triggered, prompting the merchant to manually confirm. The confirmation result is then used as a new training sample to incrementally update the model during idle periods, enabling the model to adapt to changes in the products sold at the stall (such as seasonal product replacements) and continuously improve recognition accuracy. This invention significantly improves the accuracy and response speed of product identification in farmers' market scenarios through a localized, lightweight, and dedicated model construction strategy, while reducing system deployment and maintenance costs.

[0007] The beneficial effects of this invention are: After deploying a smart traceability scale to a designated stall, it is controlled to enter learning mode, collecting image data of various products placed on the scale pan. Then, based on the product name labeling information input through the interactive interface, the image data is associated with the product name labeling information to construct a local training dataset. This local training dataset is used to train a lightweight image recognition model locally on the smart traceability scale, obtaining a dedicated image recognition model specific to the designated stall. This solution enables the smart traceability scale to train its model based on the actual product categories and appearance characteristics sold at the stall. The product recognition model is highly matched to the visual attributes such as shape, color, and texture of the specific products at that stall. In subsequent working modes, when performing real-time recognition of products on the scale pan, the accuracy of product recognition for that stall is significantly improved. Because model training is completed locally on the device, there is no need to upload product image data to the cloud, avoiding dependence on a continuous and stable network connection. High-precision product recognition can still be achieved in environments with unstable network conditions, such as farmers' markets. Local training also reduces latency caused by data transmission, accelerates model iteration and product recognition response speed, and reduces privacy risks caused by data transmission, ensuring the local processing of merchants' product sales data.

[0008] By binding the unique device identification code of the smart traceability scale with the farmers' market signage and stall identification during the device activation phase, registration and activation are completed, and a mapping table between the three is stored on a cloud server. This solution achieves a precise correspondence between the smart traceability scale and the specific business location and merchant. Subsequent training and use of the dedicated image recognition model are based on this binding relationship, ensuring that the model only serves the stall bound to the device. The mapping table recorded on the cloud server allows farmers' market managers to clearly trace the deployment location and affiliated merchant of each smart traceability scale, facilitating device management, recognition model version tracking, and transaction data collection. When devices need to be updated or replaced, the association can be quickly re-established based on the binding relationship, avoiding the problem of mismatch between the recognition model and product categories caused by chaotic device information, and improving the management efficiency and adaptability of the entire recognition system in the complex operating environment of farmers' markets. Attached Figure Description

[0009] The invention will now be further described with reference to the accompanying drawings.

[0010] Figure 1 This is a schematic diagram illustrating the working principle of a precise identification method and system based on an intelligent traceability scale as described in this invention. Figure 2 This is a flowchart of standardized commodity image acquisition in the learning mode of the intelligent traceability scale; Figure 3 This is a flowchart of the process for labeling and constructing the training dataset for product images of the intelligent traceability scale. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] See Figure 1This invention provides a precise identification method based on an intelligent traceability scale, comprising: obtaining a unique device identification code for the intelligent traceability scale; binding the unique device identification code with the agricultural market sign and stall sign to complete the registration and activation of the intelligent traceability scale; after the intelligent traceability scale is deployed to a designated stall, controlling the intelligent traceability scale to enter a learning mode, and collecting image data of various commodities placed on the weighing pan in the learning mode; associating the image data with the commodity name labeling information input through the interactive interface to construct a local training dataset; using the local training dataset to train a lightweight image recognition model locally on the intelligent traceability scale to obtain a dedicated image recognition model specific to the designated stall; switching the intelligent traceability scale from the learning mode to the working mode, and in the working mode, using the dedicated image recognition model to perform real-time identification of the commodities to be identified on the weighing pan and outputting the commodity identification result.

[0013] Example 1: In specific implementation, the device label of the smart traceability scale is scanned by the back-end management system to extract the unique device identification code. The back-end management system is equipped with a scanning module, which performs optical recognition on the barcode or QR code on the device label, and obtains the unique device identification code after parsing. The unique device identification code is an unalterable hardware identifier written into the smart traceability scale at the factory.

[0014] In practice, the system receives input information for the farmers' market name and stall number. The market name is entered by the administrator via keyboard or touchscreen on the backend management system interface, as is the stall number. The backend management system converts the market name into a market identifier, a predefined numerical code that corresponds one-to-one with the market name; it also converts the stall number into a stall identifier, also a predefined numerical code that corresponds one-to-one with the stall number. This conversion is accomplished by querying a mapping dictionary stored within the system. Each record in the mapping dictionary contains the correspondence between the market name and market identifier, as well as the correspondence between the stall number and stall identifier.

[0015] In implementation, a mapping table is established linking the unique device identification code to the farmers' market identifier and the stall identifier. This mapping table is stored as a database table within the backend management system, containing three fields: unique device identification code, farmers' market identifier, and stall identifier. Each record associates the unique device identification code, farmers' market identifier, and stall identifier, indicating that the smart traceability scale corresponding to the unique device identification code belongs to the stall corresponding to the specific farmers' market identifier. The mapping table is then stored on a cloud server, a remote storage server deployed on the internet, using a relational database management system for persistent storage.

[0016] In practice, an activation confirmation command is sent to the smart traceability scale. The backend management system sends the activation confirmation command to the smart traceability scale via a wireless network. The activation confirmation command is a message containing a specific protocol identifier, and the message content includes a unique device identification code and an activation flag. After receiving the activation confirmation command, the smart traceability scale parses the activation flag in the message, switches its status from factory default to activated, and enters the deployment-pending state. The deployment-pending state indicates that the smart traceability scale has completed registration and activation and is waiting to be deployed to the designated stall.

[0017] In practice, the association mapping table also records the version number and update timestamp of the dedicated image recognition model corresponding to each stall identifier. Two fields are added to the association mapping table: a version number field and an update timestamp field. The version number field stores the version identifier of the dedicated image recognition model, using the format of a major version number plus a minor version number, such as "1.0". The update timestamp field stores the time of the most recent update of the dedicated image recognition model, in the format of "year-month-day hour:minute:second". When the dedicated image recognition model is updated, the backend management system synchronously updates the version number and update timestamp of the corresponding stall identifier in the association mapping table.

[0018] Example 2: In specific implementation, refer to Figure 2 The system sends a mode switching command to the intelligent traceability scale. This command is triggered by the backend management system or the interactive interface on the intelligent traceability scale and received via the scale's wireless communication module. The mode switching command is a binary message containing a specific operation code, which indicates that the operating state is being switched from standby mode to learning mode. Upon receiving the mode switching command, the intelligent traceability scale parses the operation code and updates the value of its internal status register from the value corresponding to the standby state (e.g., 0x00) to the value corresponding to the learning mode (e.g., 0x01). The intelligent traceability scale then enters learning mode.

[0019] In practice, under learning mode, merchants are sequentially prompted to place each type of product in the center of the weighing pan. The prompts are displayed as text or icons on the smart traceability scale's interface, or via voice prompts played through the built-in speaker. Each product represents a category to be identified, such as a specific fruit or vegetable. Following the prompts, the merchant places one product in the center of the weighing pan, ensuring the product's main body is near the geometric center of the pan and the pan's boundaries are clearly visible.

[0020] In practice, a built-in high-definition camera captures at least one image of each product from multiple preset angles. The high-definition camera is fixed above or to the side of the smart traceability scale, with its optical axis forming a certain angle with the scale pan. The preset angles include, but are not limited to: a top-down angle, a 45-degree angle to the left, a 45-degree angle to the right, and a front-facing angle. For each product, after the merchant places it in the center of the scale pan, the camera is triggered to capture one image from each preset angle sequentially. If the number of preset angles is N, then N images are captured for each product. After all angles have been captured, a multi-angle image set for that product is obtained, containing images of that product from all angles.

[0021] In practice, for each product, each image in the multi-angle image set is aligned and cropped with the boundary of the weighing pan. The alignment and cropping process includes: using image processing algorithms to identify the outline of the weighing pan in each image (the outline can be rectangular or circular), and extracting the pixel coordinates of the weighing pan boundary using edge detection algorithms; based on the position of the weighing pan boundary, cropping the background pixels outside the weighing pan boundary, retaining the area inside the weighing pan and the product area; performing the same cropping operation on all angle images to obtain a standardized product image. The size of the standardized product image is consistent with the pixel area size within the weighing pan boundary, and the image does not contain the background outside the weighing pan.

[0022] In practice, standardized product images are cached in local storage according to their acquisition time sequence. The local storage is either a built-in non-volatile memory chip or an SD card in the smart traceability scale. The acquisition time sequence refers to the order in which multi-angle images of each product are acquired, arranged in ascending order of timestamps. Each standardized product image is saved along with its acquisition timestamp, forming an original image cache queue. The original image cache queue uses a first-in, first-out (FIFO) data structure, where each element contains the binary data of the standardized product image and its corresponding acquisition timestamp.

[0023] Example 3: In specific implementation, refer to Figure 3During the image data acquisition process, a candidate list of product names is displayed through an interactive interface. The interactive interface is a touchscreen display on the smart traceability scale. The candidate list contains multiple preset standard product name options, which are pre-stored product name sets by the system, such as "apple," "banana," and "tomato." The candidate list is displayed on the interactive interface in a vertical list or grid format, with each option displayed as a clickable button or text box.

[0024] In practice, the system receives either standard product names selected by merchants from a candidate list or custom product names manually entered by merchants. Merchants can select a standard product name option by touching a candidate list item on the interactive interface. The interface then uses that standard product name as the labeling information for the currently collected product. If no matching product name is found in the candidate list, the interface provides a text input box. Merchants can enter a custom product name using a virtual keyboard; the custom product name is a string of characters, such as "Red Fuji Apple". The interface then uses this custom product name as the labeling information for the currently collected product.

[0025] In practice, a one-to-many association is established between the product name labeling information and all standardized product images corresponding to the currently collected product. All standardized product images corresponding to the currently collected product are the standardized product images of the product from all angles obtained and cropped in the above embodiment. The association is implemented through a directory structure or database index: a folder named after the product name labeling information is created in local storage, and all standardized product image files for the product are copied or moved to this folder; or the product name labeling information field is written into the metadata of the image files. After the association is completed, labeled product image samples are generated, each labeled product image sample consisting of a standardized product image and its corresponding product name labeling information.

[0026] In practice, all labeled product image samples are compiled into a local training dataset. The local training dataset is a collection of files stored in the local storage space of the smart traceability scale, typically organized as a compressed file or database table. This local storage space is the smart traceability scale's built-in non-volatile memory. Each record in the local training dataset contains the image's binary data and the corresponding product name label string.

[0027] In practice, the candidate list of product names is dynamically generated based on the historical transaction frequency statistics of products corresponding to the farmers' market identifier. The historical transaction frequency statistics are maintained by the backend management system, which records the transaction flow of all stalls under each farmers' market identifier and calculates the frequency of each product name. For the farmers' market identifier currently bound to the smart traceability scale, the backend management system queries the historical transaction data corresponding to that identifier and sorts the product names from highest to lowest transaction frequency. The sorted product name list is then sent to the smart traceability scale via wireless network. The smart traceability scale displays the sorted product name list as a candidate list on the interactive interface. The product name with the highest frequency is displayed at the top of the candidate list, the second highest frequency is displayed second, and so on.

[0028] Example 4: In specific implementation, multiple standardized product images corresponding to each product category are extracted from the local training dataset. The local training dataset contains standardized product image files organized by product name labeling information. For each product category, i.e., each unique product name labeling information, all standardized product image files in that directory are read to obtain multiple standardized product images. Each standardized product image is converted into a pixel matrix of a preset size, set to 224×224 pixels, which is the standard input size of the lightweight convolutional neural network structure MobileNet. The conversion process includes: scaling the standardized product image to 224×224 pixels, and extracting the values ​​of the three RGB channels of each pixel to form a matrix of shape... The three-dimensional pixel matrix is ​​used as the input feature of the model.

[0029] In the specific implementation, a lightweight convolutional neural network (CNN) structure is used to build the initial recognition model on the local processor of the smart traceability scale. The lightweight CNN structure chosen is the MobileNet architecture, which consists of depthwise separable convolutional layers, batch normalization layers, ReLU activation function layers, global average pooling layers, and fully connected layers connected sequentially. The depthwise separable convolutional layers include channel-wise convolutions and pointwise convolutions. The channel-wise convolutional kernel size is 3×3 with a stride of 2, while the pointwise convolutional kernel size is 1×1 with a stride of 1. The global average pooling layer reduces the spatial dimension of the feature map to 1×1, and the number of neurons in the fully connected layers equals the total number of product categories, C. The number of convolutional layers in the lightweight CNN structure does not exceed a preset upper limit, set at 15 layers. This upper limit is determined based on the computing power and memory capacity of the local processor of the smart traceability scale, ensuring that the model can complete forward inference and backward propagation on the local processor.

[0030] In practice, the pixel matrix of each product category and its corresponding product name annotation information are used as training sample pairs. Before training, the product name annotation information is converted into a one-hot encoded vector of length C, where the category index position corresponding to the product name annotation information is 1, and the remaining positions are 0. The pixel matrix serves as the input feature, and the one-hot encoded vector represents the true label. Iterative training is performed on the initial recognition model, updating the model's weight parameters in each iteration.

[0031] In specific implementation, in the first In each iteration, a batch of training sample pairs is randomly selected from the local training dataset, and the batch size is denoted as . Batch size Setting it to 32 is a commonly used batch size in lightweight model training, balancing training speed and gradient stability within the memory limitations of the local processor. For the first... For each sample, the pixel matrix is ​​input into the current model. The model calculates the score for each category in the output layer through forward propagation. The output layer uses a fully connected layer to generate the raw score vector, which is then converted into a predicted probability vector by the Softmax function. Predicted probability vector The length is C, and each element... The model predicts the first... The sample belongs to the first The probability of each category.

[0032] In practice, the following cross-entropy loss function is used to calculate the loss value of a single sample. : in: This represents the total number of product categories. For the first The one-hot encoded vector of the true label of each sample is in the th... The values ​​(0 or 1) for each category. The model predicts the first The sample belongs to the first Calculate the probability values ​​for each category. Calculate the average loss for this batch. and according to The Adam optimization algorithm is used to update the weight parameters of the initial recognition model. The first-order moment decay coefficient of the Adam optimization algorithm is also used. Set to 0.9, second-order moment attenuation coefficient The values ​​are set to 0.999 and the learning rate is set to 0.001; these are the standard recommended values ​​for the Adam optimization algorithm.

[0033] In practice, iterative training stops when the recognition accuracy of the training sample pairs reaches a preset convergence threshold. The preset convergence threshold is set to 0.95, meaning that the model is considered converged when its classification accuracy on the local training dataset is not lower than 95%. The current model parameters are then fixed to obtain a dedicated image recognition model. The model parameters include the weight matrices and bias vectors of all convolutional and fully connected layers, which are stored as binary files in the local storage space of the smart traceability scale.

[0034] In practical implementation, the lightweight image recognition model adopts MobileNet, ShuffleNet, or EfficientNet-Lite architectures. During model training, the input image undergoes random rotation, scaling, and color dithering enhancement. The random rotation angle range is set to [-15°, 15°], the random scaling ratio range is set to [0.9, 1.1], and the color dithering includes brightness adjustment (coefficient range [0.8, 1.2]), contrast adjustment (coefficient range [0.8, 1.2]), and saturation adjustment (coefficient range [0.8, 1.2]). These enhancements are randomly applied to the current batch of pixel matrices in each iteration before being input into the model to increase the diversity of the training data.

[0035] Example 5: In specific implementation, upon receiving the mode switching completion instruction, the operating status of the smart traceability scale is marked as working mode. The mode switching completion instruction is triggered by the merchant clicking the "Start Working" button through the interactive interface. After receiving the instruction, the operating system of the smart traceability scale updates the internal status register from the value corresponding to the learning mode (e.g., 0x01) to the value corresponding to the working mode (e.g., 0x02).

[0036] In practical implementation, under working mode, when a product to be identified is detected placed on the weighing pan, the camera is triggered to capture an image of the product. The detection method involves a pressure sensor beneath the weighing pan monitoring weight changes in real time. When the weight changes from zero to non-zero and remains stable for more than 0.5 seconds, it is determined that a product has been placed, and a trigger signal is sent to the camera. Upon receiving the trigger signal, the camera captures an image with a resolution of 1920×1080 pixels. The current image undergoes denoising and size normalization. Denoising uses a median filter algorithm with a filtering window size of 3×3 pixels. Size normalization scales the image to 224×224 pixels to obtain the image to be identified, which is an RGB three-channel color image.

[0037] In practice, the image to be recognized is input into a dedicated image recognition model, which then outputs a probability distribution vector indicating which category the product belongs to. The specific process includes: converting the image to be recognized into a format with a size of [insert size here]. pixel matrix ,in and The output resolution remains consistent with that mentioned in the above embodiment, i.e. , Pixel matrix The shape is Each element is an integer value between 0 and 255. The input is a dedicated image recognition model, which calculates the original score vector of the output layer through forward propagation. ,in This represents the total number of product categories. For the first The raw scores for each category are output from the last fully connected layer of the model, without an activation function. The following Softmax function is used to... Convert to probability distribution vector : in: is the base of the natural logarithm, with a value of approximately 2.71828. This indicates that the dedicated image recognition model determines that the product to be identified belongs to the category of [missing information]. The probability values ​​of each category, and satisfying Probability distribution vector Each element in The value ranges from 0 to 1, and the sum of all elements equals 1.

[0038] In practical implementation, the probability distribution vector The output result is passed to the product recognition result output module of the smart traceability scale. This module is a software module running on the processor of the smart traceability scale. It selects the product name corresponding to the category with the highest probability value from the probability distribution vector as the product recognition result output. Specifically, it uses the argmax function to find the index. According to the index The system searches for the corresponding product name string in the product name mapping table and outputs the string to the display interface and pricing module of the intelligent traceability scale.

[0039] In practice, after selecting the product name corresponding to the category with the highest probability value from the probability distribution vector as the product identification result, the process also includes a confidence check of the identification result. The highest probability value in the probability distribution vector is extracted as the first confidence level. The second highest probability value is extracted as the second confidence level. Calculate the difference between the first confidence level and the second confidence level. The difference The result is compared with a preset confidence threshold, which is set to 0.3. This threshold is determined based on the model's validation results on the training set. A difference of 0.3 or greater is considered a reliable recognition result. When the difference is greater than or equal to the confidence threshold (0.3), the product identification result is directly output to the pricing module. When the confidence level is less than the threshold (0.3), a vague identification signal is generated, and the merchant is prompted to manually confirm the product to be identified via the interactive interface. The vague identification signal is a Boolean variable with a value of true; the prompt is displayed as a pop-up window on the interactive interface, containing the message "Vague identification, please manually select a product name" and a list of candidate product names.

[0040] In practical implementation, when the difference is less than the confidence threshold, a recognition ambiguity signal is generated, and after prompting the merchant to manually confirm the product to be identified through the interactive interface, the process also includes an incremental update step for the dedicated image recognition model. The system receives the manually confirmed product name input by the merchant through the interactive interface. The manually confirmed product name can be a standard product name selected by the merchant from the candidate list or a custom product name manually entered. This manually confirmed product name is used as the real label. The image to be identified corresponding to the product to be identified is combined with the real label to form a new training sample. The new training sample consists of a product image to be identified (224×224 pixel matrix) and a string label. The new training sample is added to the local training dataset, which is stored in the local storage space of the smart traceability scale. The method of adding is to save the image file to be identified into a folder named after the manually confirmed product name, and update the dataset index file. This yields an expanded local training dataset. During the idle periods of the smart traceability scale, incremental training is performed on the dedicated image recognition model using the expanded local training dataset, updating the model parameters of the dedicated image recognition model. The idle period is defined as the time during which the intelligent traceability scale does not detect any goods on the weighing pan and does not receive any operation commands; it enters the idle state if the duration exceeds 10 minutes. Incremental training uses the same lightweight convolutional neural network structure and training parameters as in the previous embodiment, only fine-tuning the existing model parameters. The learning rate is adjusted to 0.0001, the number of training epochs is set to 5, and the batch size... Set to 16. The updated model parameters will overwrite the original dedicated image recognition model parameter file.

[0041] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A precise identification method based on an intelligent traceability scale, characterized in that, The method includes: Obtain the unique device identification code of the smart traceability scale, and bind the unique device identification code with the farmers' market sign and stall sign to complete the registration and activation of the smart traceability scale; After the smart traceability scale is deployed to the designated stall, the smart traceability scale is controlled to enter the learning mode, and image data of various products placed on the weighing pan are collected in the learning mode; Based on the product name labeling information input through the interactive interface, the image data is associated with the product name labeling information to construct a local training dataset; Using the local training dataset, a lightweight image recognition model is trained locally on the smart traceability scale to obtain a dedicated image recognition model specific to the designated stall, specifically including: Multiple standardized product images corresponding to each product category are extracted from the local training dataset, and each standardized product image is converted into a pixel matrix of a preset size as the input features of the model. An initial recognition model is built on the local processor of the smart traceability scale using a lightweight convolutional neural network structure, wherein the number of convolutional layers in the lightweight convolutional neural network structure does not exceed a preset upper limit. The pixel matrix of each type of product and its corresponding product name labeling information are used as training sample pairs to perform iterative training on the initial recognition model, and the weight parameters of the model are updated in each iteration. When the recognition accuracy of the training sample pairs reaches the preset convergence threshold, the iterative training stops, the current model parameters are fixed, and the exclusive image recognition model is obtained. The intelligent traceability scale is switched from the learning mode to the working mode. In the working mode, the dedicated image recognition model is used to identify the product to be identified on the scale pan in real time and output the product identification result.

2. The method for accurate identification based on an intelligent traceability scale according to claim 1, characterized in that, The steps of obtaining the unique device identification code of the smart traceability scale, binding the unique device identification code with the farmers' market signage and stall signage, and completing the registration and activation of the smart traceability scale specifically include: The unique device identification code is extracted by scanning the device label of the smart traceability scale through the background management system. Receive input farmers' market name information and stall number information, convert the farmers' market name information into the farmers' market identifier, and convert the stall number information into the stall identifier; Establish a mapping table between the unique device identification code, the farmers' market identifier, and the stall identifier, and store the mapping table in a cloud server; Send an activation confirmation command to the smart traceability scale to enable the smart traceability scale to complete registration and activation and enter the deployment state.

3. The method for accurate identification based on an intelligent traceability scale according to claim 2, characterized in that, The association mapping table also records the version number and update timestamp of the exclusive image recognition model corresponding to each stall identifier.

4. The method for accurate identification based on an intelligent traceability scale according to claim 1, characterized in that, After the intelligent traceability scale is deployed to a designated stall, the steps of controlling the intelligent traceability scale to enter learning mode and collecting image data of various products placed on the weighing pan in learning mode specifically include: Send a mode switching command to the intelligent traceability scale to switch the intelligent traceability scale from standby mode to learning mode; In the learning mode, the merchant is prompted to place each product in the center of the weighing pan, and the built-in high-definition camera captures at least one image of each product from multiple preset angles to obtain a multi-angle image set of each product. For each product, each image in the multi-angle image set is aligned with the boundary of the weighing pan and cropped to remove the background area outside the weighing pan, resulting in a standardized product image. The standardized product images are cached in local storage according to the acquisition time sequence to form an original image cache queue.

5. The method for accurate identification based on an intelligent traceability scale according to claim 1, characterized in that, The step of associating the image data with the product name labeling information input through the interactive interface to construct a local training dataset specifically includes: During the process of acquiring the image data, a candidate list of product names is displayed through the interactive interface. The candidate list of product names contains multiple preset standard product name options. The system receives standard product names selected by merchants from the product name candidate list, or custom product names manually entered by merchants, and uses the received names as the product name labeling information for the currently collected products. Establish a one-to-many association between the product name labeling information and all standardized product images corresponding to the currently collected product, and generate labeled product image samples; All labeled product image samples are aggregated into the local training dataset, and the local training dataset is stored in the local storage space of the smart traceability scale.

6. The method for accurate identification based on an intelligent traceability scale according to claim 5, characterized in that, The candidate list of product names is dynamically generated based on the statistical results of the historical transaction frequency of the products corresponding to the farmers' market identifier.

7. The method for accurate identification based on an intelligent traceability scale according to claim 1, characterized in that, The lightweight image recognition model adopts MobileNet, ShuffleNet, or EfficientNet-Lite architecture, and performs random rotation, scaling, and color dithering enhancement on the input image during model training.

8. The method for accurate identification based on an intelligent traceability scale according to claim 1, characterized in that, Switching the intelligent traceability scale from the learning mode to the working mode, and in the working mode, using the dedicated image recognition model to identify the product to be identified on the scale pan in real time, and outputting the product identification result, specifically includes the following steps: Upon receiving the mode switching completion instruction, the operating status of the intelligent traceability scale is marked as the working mode; In the operating mode, when a product to be identified is detected placed on the weighing pan, the camera is triggered to capture the current image of the product to be identified, and the current image is processed for noise reduction and size normalization to obtain the image to be identified; The image to be identified is input into the dedicated image recognition model, and the dedicated image recognition model outputs the probability distribution vector of the product to be identified belonging to each category. The product name corresponding to the category with the highest probability value is selected from the probability distribution vector and output as the product identification result.

9. A precision identification system based on an intelligent traceability scale, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the precise identification method based on the intelligent traceability scale as described in any one of claims 1 to 8.

Images