Recyclate identification processing method, apparatus, device, and medium

Abstract

This application provides a method, apparatus, device, and medium for identifying and processing recyclable materials. In this method, image data and weight data of the item to be recycled are acquired; the image data are input into a first recognition model and a second recognition model respectively, and inference is performed in parallel to output corresponding material identification results and product identification results; based on the material identification results, product identification results, and weight data, it is determined whether the recyclable item is a qualified recyclable; if it is determined to be a qualified recyclable, a first control command is output, or if it is determined to be an unqualified recyclable, a second control command is output. This solution addresses the problem in related technologies where recyclable material identification schemes cannot meet the requirement of acquiring directly recyclable aluminum cans and determining their product attributes, thus affecting the recycling effect.

Description

Technical Field

[0001] This application relates to the field of resource recycling technology, and in particular to a method, apparatus, equipment and medium for identifying and processing recyclables. Background Technology

[0002] With increasing emphasis on resource recycling, the recycling of waste packaging materials, especially aluminum beverage cans, has significant economic and environmental value. Since the energy consumption for reprocessing recycled aluminum is only 5% or even less than that of electrolytic aluminum production, it can significantly reduce production costs. Therefore, targeted recycling of aluminum cans has become a research direction in the resource regeneration industry. However, current mainstream waste recycling methods mostly employ incineration, landfill, or mixed sorting, making it difficult to accurately screen specific materials (such as pure aluminum). Even in the field of intelligent recycling equipment, existing technologies mainly remain at a coarse-grained level of waste sorting (such as distinguishing between recyclable and non-recyclable materials), lacking the ability to fine-grainedly identify the purity of recycled materials (such as distinguishing between pure aluminum and aluminum alloys) and product attributes (such as beverage brands). Furthermore, existing algorithms often struggle to effectively assess whether items meet the conditions for direct recycling (such as the presence of residual liquid or severe deformation of the can), resulting in inconsistent quality of recycled materials and impacting the economic benefits of high-value recycling.

[0003] To address the aforementioned issues, vision-based recyclable identification solutions exist in related technologies. However, these methods often employ serial processing pipelines or single-modal recognition, making it difficult to achieve high-precision, multi-dimensional real-time identification simultaneously with limited computing resources. This results in an inability to meet the requirement of acquiring directly recyclable aluminum cans and determining their commodity attributes, thus affecting the recycling effect. Summary of the Invention

[0004] The recyclable identification and processing method, apparatus, equipment, and medium provided in this application are intended to solve the problem that recyclable identification schemes in related technologies cannot meet the requirements of obtaining aluminum cans that can be directly recycled and determining their commodity attributes, thus affecting the recycling effect.

[0005] In a first aspect, embodiments of this application provide a method for identifying and processing recyclable materials, applied in the processor of a recycling machine, comprising:

[0006] Obtain image and weight data of the items to be recycled;

[0007] Image data is input into the first recognition model and the second recognition model respectively, and inference is performed in parallel to output the corresponding material recognition result and product identification result; wherein, the material recognition result is output by the first recognition model and the product identification result is output by the second recognition model. The first recognition model and the second recognition model are models that are deployed independently in the processor after being lightweighted.

[0008] Based on the material identification results, product label identification results, and weight data, determine whether the recycled items are qualified recyclables;

[0009] If the item is determined to be a qualified recyclable, a first control command is output; or if the item is determined to be an unqualified recyclable, a second control command is output. The first control command instructs the conveyor to transport the item to be recycled to the collection bin and triggers a reward feedback operation. The second control command instructs the conveyor to return the item to be recycled to the exit.

[0010] In one embodiment of this disclosure, the first identification model is a pre-trained model for identifying the material category of the items to be recycled, and the second identification model is a pre-trained model for identifying the product label category of the items to be recycled; the first identification model and the second identification model are stored as independent model files, and the second identification model supports iterative updates independently of the first identification model to adapt to newly added product label categories.

[0011] In one embodiment of this disclosure, lightweight processing is achieved as follows: during the training phase, a data augmentation strategy is used to simulate the flattening, occlusion, and lighting changes in the recycling scene to train the first recognition model and the second recognition model; after training, the trained floating-point model is converted into an INT8 quantization model and calibrated using a calibration set containing real scene samples until the accuracy loss of the quantization model is lower than a preset value.

[0012] In one embodiment of this disclosure, acquiring image data of the item to be recycled includes: continuously reading video frames from a camera via a background thread and updating the latest frame to a shared memory buffer in real time; and reading the latest frame from the shared memory buffer as image data in response to a trigger signal generated when the recycled item reaches a preset identification position.

[0013] In one embodiment of this disclosure, image data is input into a first recognition model and a second recognition model respectively, and inference is performed in parallel to output the corresponding material recognition result and product identification result. This includes: running the inference tasks corresponding to the first recognition model and the second recognition model respectively in the processor based on different threads in multi-threading technology; and outputting the material recognition result and product identification result after the first recognition model and the second recognition model have completed inference based on the thread synchronization mechanism.

[0014] In one embodiment of this disclosure, determining whether a recyclable item is a qualified recyclable based on material identification results, product identification results, and weight data includes: obtaining a first confidence level of the material identification results and a second confidence level of the product identification results; determining a corresponding preset weight range based on the material identification results and / or product identification results; if the first confidence level is greater than a first preset threshold, the second confidence level is greater than a second preset threshold, and the weight data is within the preset weight range, then the item to be recycled is determined to be a qualified recyclable; otherwise, it is determined to be an unqualified recyclable.

[0015] In one embodiment of this disclosure, after acquiring image data and weight data of the item to be recycled, the method further includes at least one of the following: performing occlusion detection based on the image data, and outputting a second control command when the occlusion rate of the key area of ​​the item to be recycled exceeds a preset occlusion threshold; performing contour analysis based on the image data, and performing a downgrade operation on the item to be recycled when the flattening rate of the item to be recycled exceeds a preset flattening threshold; matching the image data with a pre-stored hazardous materials database, and performing double verification in conjunction with abnormal weight data; if the match is successful and the weight exceeds a preset normal range, the item to be recycled is identified as a hazardous material and an alarm or return operation is triggered.

[0016] Secondly, embodiments of this application provide a recyclable material identification and processing device, applied in the processor of a recycling machine, comprising:

[0017] The acquisition module is used to acquire image data and weight data of the items to be recycled;

[0018] The recognition module is used to input image data into the first recognition model and the second recognition model respectively, perform inference in parallel, and output the corresponding material recognition result and product identification result; wherein, the material recognition result is output by the first recognition model and the product identification result is output by the second recognition model. The first recognition model and the second recognition model are models that are deployed independently in the processor after being lightweighted.

[0019] The analysis module is used to determine whether the recycled items are qualified recyclables based on the material identification results, product label identification results, and weight data.

[0020] The processing module is used to output a first control command if the item is determined to be a qualified recyclable, or a second control command if the item is determined to be an unqualified recyclable. The first control command is used to instruct the conveying mechanism to convey the item to be recycled to the collection bin and trigger a reward feedback operation. The second control command is used to instruct the conveying mechanism to return the item to be recycled to the exit.

[0021] Optionally, the recognition module is specifically used for: a first recognition model is a pre-trained model for recognizing the material category of the item to be recycled, and a second recognition model is a pre-trained model for recognizing the commodity label category of the item to be recycled; the first recognition model and the second recognition model are stored as independent model files, and the second recognition model supports iterative updates independently of the first recognition model to adapt to newly added commodity label categories.

[0022] Optionally, the recognition module is specifically used to achieve lightweight processing in the following way: during the training phase, a data augmentation strategy is used to simulate the flattening, occlusion and lighting changes in the recycling scene to train the first recognition model and the second recognition model; after training, the trained floating-point model is converted into an INT8 quantization model and calibrated using a calibration set containing real scene samples until the accuracy loss of the quantization model is lower than a preset value.

[0023] Optionally, the acquisition module is specifically used to continuously read video frames from the camera through a background thread and update the latest frame to the shared memory buffer in real time; in response to the trigger signal generated when the recycled item reaches the preset recognition position, it reads the latest frame from the shared memory buffer as image data.

[0024] Optionally, the recognition module is specifically used to run the inference tasks corresponding to the first recognition model and the second recognition model respectively in the processor based on different threads in multi-threading technology; and to output the material recognition result and the product identification result after the first recognition model and the second recognition model have completed inference based on the thread synchronization mechanism.

[0025] Optionally, the analysis module is specifically used to: obtain a first confidence level of the material identification result and a second confidence level of the product identification result; determine the corresponding preset weight range based on the material identification result and / or the product identification result; if the first confidence level is greater than a first preset threshold, the second confidence level is greater than a second preset threshold, and the weight data is within the preset weight range, then the item to be recycled is determined to be a qualified recyclable; otherwise, it is determined to be an unqualified recyclable.

[0026] Optionally, the processing module is further configured to, after acquiring the image data and weight data of the item to be recycled, perform at least one of the following actions: perform occlusion detection based on the image data, and output a second control command when the occlusion rate of the key area of ​​the item to be recycled exceeds a preset occlusion threshold; perform contour analysis based on the image data, and perform a downgrade operation on the item to be recycled when the flattening rate of the item to be recycled exceeds a preset flattening threshold; match the image data with a pre-stored hazardous materials database, and perform double verification in conjunction with abnormal weight data; if the match is successful and the weight exceeds a preset normal range, then the item to be recycled is identified as a hazardous material and an alarm or return operation is triggered.

[0027] Thirdly, embodiments of this application provide an electronic device, including:

[0028] Processor, memory, communication interface;

[0029] Memory is used to store the processor's executable instructions;

[0030] The processor is configured to execute the recyclable identification processing method of any one of the first aspects by executing executable instructions.

[0031] Fourthly, embodiments of this application provide a readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the recyclable identification and processing method of any one of the first aspects.

[0032] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, is used to implement the recyclable identification and processing method of any one of the first aspects.

[0033] The recyclable material identification and processing method, apparatus, device, and medium provided in this application acquire image data and weight data of the items to be recycled. The image data is then input into a first and a second identification model, which are independently deployed after lightweight processing, and inference is performed in parallel. Corresponding material identification results and product identification results are output. Based on the fusion of the material identification results, product identification results, and weight data, a determination is made as to whether the item is a qualified recyclable, and corresponding control commands are output. This effectively solves the technical problems of model coupling leading to update difficulties and serial identification failing to meet real-time requirements in existing recycling identification schemes. It achieves a technological breakthrough by simultaneously performing high-precision material identification and product identification on edge computing devices, significantly improving the accuracy and processing efficiency of fine-grained identification of recyclable items, while reducing model maintenance costs and hardware deployment barriers. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0035] Figure 1 This application provides a schematic diagram illustrating the working principle of the recycling machine.

[0036] Figure 2 This is a flowchart illustrating an embodiment of the recyclable material identification and processing method provided in this application.

[0037] Figure 3 This is a flowchart illustrating an embodiment of the recyclable material identification and processing method provided in this application.

[0038] Figure 4 A schematic diagram of the structure of an embodiment of the recyclable material identification and processing device provided in another embodiment of this application;

[0039] Figure 5 This is a schematic diagram of the structure of an electronic device provided in this application.

[0040] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0041] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0042] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0043] Specific material recycling, exemplified by aluminum recycling, holds significant economic value in the resource regeneration sector. The energy consumption for reprocessing recycled aluminum is far lower than that for producing primary aluminum, making the accurate screening of high-value recyclables such as aluminum cans a key industry requirement. However, in actual recycling scenarios, it's not only necessary to distinguish between aluminum and non-aluminum materials, but also to identify material purity and product labeling. This allows for differentiated recycling and reward feedback based on brand or model, thereby increasing user participation in aluminum recycling activities and improving recycling efficiency. Therefore, the identification system must possess multi-dimensional, fine-grained recognition capabilities.

[0044] Most existing recycling identification solutions remain at a coarse-grained level of waste sorting, failing to meet the demand for precise identification of specific material purity and product labels. Some vision-based identification solutions employ a single model or serial processing architecture, which struggles to balance recognition accuracy and processing speed under limited edge computing resources. While solutions using large neural network models offer higher accuracy, they rely on high-performance hardware such as graphics processing units, resulting in high costs and hindering large-scale deployment. Furthermore, existing solutions do not adequately consider model updates and maintenance; adding new product labels often requires retraining the entire model, leading to high operating costs.

[0045] The recyclable material identification and processing method provided in this application decouples material identification and product identification into two independent lightweight models, which are deployed in parallel within the recycling machine's processor. After acquiring image and weight data, dual-model inference is executed synchronously, and the judgment is made based on the fusion of multi-dimensional identification results and weight data. Thus, the decoupled parallel architecture resolves the inherent contradiction between model update and maintenance and real-time processing efficiency, achieving efficient operation at the edge while ensuring fine-grained identification accuracy. This overcomes the major obstacle of existing technologies that struggle to balance accuracy, efficiency, and maintainability.

[0046] Figure 1 This is a schematic diagram illustrating the working principle of a recycling machine for the application of the recycling identification and processing method provided in this application. During the recycling identification and processing process, the recycling machine 100 receives items deposited by the user through the delivery port 101, and transfers the items to be recycled to the rear through the conveying mechanism 102. The weight is measured by the weighing sensor 103, and the position sensor 104 measures whether the item has moved to a position where an image can be captured. Then, the image data is captured by the camera 105, and the weight data is combined for identification. After confirming that the item can be recycled, it is moved to the collection bin 106 through the conveying mechanism 102.

[0047] It should be noted that, Figure 1 The scenario shown includes a recycling machine, a delivery port, a conveying mechanism, a weighing sensor, a position sensor, a camera, and a storage compartment. Only one or a specific number of these are used as examples for illustration, but this disclosure is not limited to this. That is to say, the number of recycling machines, delivery ports, conveying mechanisms, weighing sensors, position sensors, cameras, and storage compartments can be arbitrary.

[0048] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0049] Figure 2 A flowchart illustrating a recyclable material identification and processing method provided in this application is shown below.Figure 2 As shown, the method includes:

[0050] S201. Obtain image data and weight data of the items to be recycled.

[0051] Specifically, this embodiment is used to illustrate the main steps of the recyclable material identification and processing method.

[0052] The execution entity in this embodiment is a processor in a recycling machine specifically designed for recycling specific materials such as aluminum. These recycling machines are typically installed in public places such as communities and supermarkets, and use incentive mechanisms to guide users to dispose of waste packaging, especially high-value recyclables such as aluminum beverage cans.

[0053] Recyclable items enter the internal conveyor mechanism of the recycling machine through the delivery port. The conveyor mechanism usually uses a crawler or roller conveyor to smoothly transport the items to the identification area.

[0054] Weighing sensors are installed below the conveyor or at a specific location to measure the weight of the items to be recycled in real time.

[0055] The weighing sensor can be either a resistance strain gauge or a capacitive sensor. When an item is placed on the conveyor belt, the sensor converts the weight into an electrical signal, which is then converted from analog to digital to obtain a digital weight value. To accommodate recyclable items of different sizes, the weighing sensor's range is typically set from 0 to 500 grams, with a measurement accuracy of ±1 gram.

[0056] Meanwhile, high-definition cameras are deployed above or to the side of the identification area to collect image data of the items to be recycled. These cameras can employ rolling shutter or global shutter CMOS sensors with a resolution of at least five megapixels to ensure clear capture of detailed features on the item's surface, including material texture, printed patterns, and brand logos.

[0057] Communication between the camera and the processor can be achieved through a universal serial bus or Ethernet interface, and the transmission bandwidth needs to meet the requirements of video stream transmission of more than 30 frames per second.

[0058] Position sensors, such as photoelectric switches, infrared sensors, or proximity sensors, are placed at key locations within the recognition area to detect whether an item has reached the preset recognition position.

[0059] When the sensor detects that an object is in place, it immediately sends a trigger signal to the processor, which then executes the image acquisition command to obtain the image frame of the current moment from the camera.

[0060] To improve the real-time performance and accuracy of image acquisition, a triggered acquisition mode can be adopted, which means that acquisition is only performed when a trigger signal is received, thus avoiding resource waste and delay caused by continuous acquisition.

[0061] In applications with poor lighting conditions, supplementary lights, such as LED ring lights or infrared supplementary lights, can be placed around the camera to ensure uniform image brightness and clear details.

[0062] In this way, the processor can simultaneously acquire image data and weight data of the items to be recycled, laying the data foundation for subsequent identification and processing.

[0063] Furthermore, regarding the collection of weight data, since the recycled items may experience slight vibrations during transportation, in order to obtain stable weight readings, a method of averaging multiple samples can be used. For example, after sampling ten times consecutively, the arithmetic mean can be calculated, and obviously abnormal sample values ​​can be removed before averaging.

[0064] At this point, the sampling frequency can be set to 100 Hz, with a sampling interval of 10 milliseconds, and the total sampling time can be controlled within 100 milliseconds to balance measurement accuracy and processing speed.

[0065] In some embodiments, the weighing sensor requires periodic calibration. Calibration can be performed automatically using a built-in standard weight or manually on a periodic basis. The calibration cycle can be set monthly or quarterly depending on the frequency of use. During calibration, the processor records the zero-point output under no-load conditions and the full-scale output after loading the standard weight, establishing a linear mapping relationship between the weight value and the sensor output signal for subsequent conversion calculations in actual measurements.

[0066] Furthermore, regarding image data acquisition, after receiving the image data, the processor can perform necessary preprocessing operations, including image cropping, scaling, rotation correction, and format conversion. Image cropping can remove background interference areas and extract the image of the object within the recognition area; scaling can uniformly adjust the image to a preset size to meet the input requirements of the subsequent recognition model; rotation correction can correct image tilt caused by incorrect object placement angles, which can be achieved by using Hough transform to detect the straight lines of the object's edges and calculate the tilt angle, followed by affine transformation correction; format conversion can convert the original image format output by the camera, such as the Bayer array format, to a red-green-blue color space format for easier subsequent processing.

[0067] S202. Input the image data into the first recognition model and the second recognition model respectively, perform inference in parallel, and output the corresponding material recognition result and product identification result.

[0068] The material recognition result is output by the first recognition model, and the product identification result is output by the second recognition model. The first and second recognition models are lightweight models that are independently deployed in the processor.

[0069] Specifically, the first identification model is a neural network model specifically trained to identify the material categories of items to be recycled. The material categories it can identify include pure aluminum, aluminum alloy, iron, plastic, paper, and other materials.

[0070] The second identification model is a neural network model specifically trained to identify the product identification categories of items to be recycled. The product identification categories it can identify include fine-grained information such as beverage brands, product series, and packaging specifications.

[0071] Both models have been lightweighted to ensure efficient operation on recycling processors. Lightweighting includes model structure compression and parameter simplification. For example, depthwise separable convolutions are used instead of standard convolutions to reduce computation and parameter count; global average pooling is used instead of fully connected layers to reduce model size; and channel pruning is used to remove redundant feature channels, further compressing the model volume.

[0072] After lightweighting, the number of parameters in the first and second recognition models is significantly reduced, enabling the models to run in real time on the processor deployed in the recycling machine without relying on dedicated hardware such as graphics processors.

[0073] Regarding model deployment, the first and second recognition models are stored as independent files on the processor's storage medium. The model file format can adopt the Open Neural Network Exchange Format, which has good cross-platform compatibility and can run on various hardware platforms and operating systems. When the processor loads a model, it reads both model files into memory and allocates an independent inference engine instance for each model. The inference engine can use a lightweight runtime framework optimized for the central processing unit, enabling efficient model computation.

[0074] During inference, the processor simultaneously transmits the preprocessed image data to the inference engines of both models, which then perform forward propagation computations separately. Since the two models are independent, their inference tasks can be executed in parallel, fully utilizing the processor's multi-core parallel computing capabilities.

[0075] In practice, an independent inference thread can be created for each model, with two threads running concurrently to perform operations such as model loading, input data preparation, forward computation, and result parsing, respectively. Thread creation and management can be implemented through a system-provided thread library, such as a portable operating system interface thread library.

[0076] The processor allocates thread priorities and scheduling strategies reasonably based on the number of cores to ensure that the two inference threads receive a fair allocation of computing resources.

[0077] After both threads have completed their inference tasks, the main thread obtains the recognition results output by the two models, including the material category and its confidence level, and the product identifier category and its confidence level.

[0078] S203. Based on the material identification results, product label identification results, and weight data, determine whether the recycled items are qualified recyclables.

[0079] Specifically, after obtaining the material identification results and the product identification results, the processor combines the weight data collected by the weighing sensor to execute multi-dimensional fusion judgment logic.

[0080] The core of the judgment logic is to assess whether an item meets the various conditions for being a qualified recyclable based on preset rules.

[0081] For material identification results, the processor focuses on whether the identified material category is the target recycled material, such as whether it is pure aluminum or a specific alloy. Material identification results are usually accompanied by a confidence score, indicating the model's degree of certainty regarding the identification result.

[0082] For product label recognition results, the processor focuses on whether the identified product label is within the recyclable range, such as whether it belongs to a beverage can from a brand that partners with a recycling program. Similarly, the product label recognition results also include a confidence score.

[0083] For weight data, the processor needs to determine whether the weight of the current item is within a reasonable range. This reasonable range can be determined based on historical data statistics or dynamically adjusted based on material category and product labeling.

[0084] In practice, a mapping table between material categories and standard weights can be established, such as the standard weight range of a 330ml pure aluminum beverage can, the standard weight range of a 500ml pure aluminum beverage can, and the standard weight range of an iron can.

[0085] The processor looks up the corresponding preset weight range in a table based on the material identification results, and then compares the actual measured weight data with the range. If the weight exceeds the range, there may be residual liquid, foreign objects, or the item may be severely deformed.

[0086] For product identification results, different brands or packaging specifications may correspond to different standard weights. The processor can further refine the determination of the weight range, for example, by selecting a more accurate weight mapping relationship based on the product identification results.

[0087] In addition, the processor can make a comprehensive judgment by combining the confidence levels of material recognition results and product identification results, and set confidence thresholds. For example, the confidence level of material recognition must be greater than 95%, and the confidence level of product identification recognition must be greater than 90%. Only recognition results that meet the confidence level requirements at the same time are considered reliable.

[0088] If the identification results are unreliable or the item does not meet the recycling conditions, the processor may determine it as an unqualified recyclable. Unqualified recyclables include, but are not limited to, the following situations: the material identification result shows that it is not the target recycling material, such as plastic bottles or paper packaging; the product label identification result shows that it is not a partner brand or the brand cannot be identified; the weight data exceeds the preset range, which may indicate the presence of residual liquid or foreign matter; the confidence of material identification or product label identification is too low, and the item's attributes cannot be determined; the item has obvious damage, deformation, or obstruction, affecting the accuracy of identification and recycling value.

[0089] In addition to meeting the above conditions, additional conditions can be set for determining whether a product is eligible for recycling, such as accepting only items within a specific size range or only accepting items that are not severely crushed.

[0090] The judgment logic can be flexibly configured according to different scenarios. The processor can adjust the judgment rules in real time by reading the configuration file or receiving the policy parameters sent from the cloud to adapt to the changing recycling requirements at different stages.

[0091] S204. If the product is determined to be a qualified recyclable, output the first control command; or if the product is determined to be an unqualified recyclable, output the second control command.

[0092] The first control command instructs the conveying mechanism to transport the items to be recycled to the collection bin and triggers a reward feedback operation; the second control command instructs the conveying mechanism to return the items to be recycled to the exit.

[0093] Specifically, after the processor completes the judgment of qualified recyclables, it generates corresponding control commands based on the judgment results and sends them to the actuator of the recycling machine through the input / output interface.

[0094] If the item is determined to be a qualified recyclable, the processor generates a first control command, which instructs the conveyor mechanism to transport the item to the collection bin. The conveyor mechanism is typically driven by a stepper motor or a DC motor, and the processor controls the motor's direction and speed via a motor driver.

[0095] When an item is in the recognition area, the conveyor is paused and waits for the recognition to complete. After the recognition is completed, if the item is determined to be a qualified product, the processor sends a forward rotation signal to the motor driver, which drives the conveyor belt to move forward and transport the item to the entrance of the storage compartment. The item falls into the storage compartment by gravity.

[0096] A dedicated detection sensor can be installed inside the storage compartment. When the storage compartment is about to be full, it sends a signal to the processor, which can trigger an alarm or notify maintenance personnel to clean it.

[0097] Simultaneously, the first control command also triggers a reward feedback operation. This reward feedback operation aims to incentivize user participation in recycling and can be implemented in various ways. One method is to generate points data and accumulate these points in the user's account. Users typically complete identification by scanning a QR code or entering their mobile phone number before submitting their recyclables. The processor can then associate and store the user's identifier with the points data. Once a certain amount of points is accumulated, they can be redeemed for physical goods or coupons.

[0098] Another approach is to directly issue electronic coupons or prize draw opportunities, which users can use in linked mini-programs or applications. The specific form of the reward feedback operation can be configured according to the operational strategy, and the processor implements it by calling external service interfaces or updating the local database.

[0099] To ensure the accuracy of reward distribution and prevent fraud, the processor can record detailed information such as the time of each qualified delivery, item identification result, and weight, and generate a unique transaction serial number for subsequent reconciliation and auditing.

[0100] If the item is determined to be an unqualified recyclable, the processor generates a second control command, which instructs the conveyor to return the item to the exit point. Upon receiving a reversal signal, the conveyor moves in the opposite direction, sending the item back from the identification area to the exit point. The exit point is usually located near the delivery port or on the other side, allowing the user to easily retrieve the returned item.

[0101] After the return operation is completed, the processor can inform the user of the reason for the return through the display screen or voice prompt, such as "non-recyclable material", "too much residual liquid", "brand not within the recycling scope", etc., to help users understand the rules and improve their subsequent delivery behavior.

[0102] In cases of multiple rejections, the processor can record user submission behavior data for subsequent analysis and rule optimization.

[0103] In some embodiments, the processor also needs to coordinate the collaborative work of various modules for the overall workflow of the recycling machine. In standby mode, the processor continuously monitors the drop-in port sensor. When it detects that a user has dropped an item, it activates the conveyor mechanism to transport the item to the recognition area and triggers image acquisition and weight measurement.

[0104] After identification, the system controls the transmission mechanism to perform corresponding actions based on the judgment result, while simultaneously updating user points and recording transaction logs. The timing control of the entire process must be precise and reliable to ensure a smooth user experience. The processor can also communicate with the cloud server via a wireless network module to upload recycled data, receive policy updates, and software upgrades.

[0105] The cloud server can aggregate data from multiple recycling machines for statistical analysis, operational optimization, and anomaly monitoring, forming a complete closed loop for recycling management.

[0106] The recyclable material identification and processing method provided in this application acquires image data and weight data of the item to be recycled. The image data is then input into a first and a second identification model, which are independently deployed after lightweight processing, and inference is performed in parallel. Corresponding material identification results and product identification results are output. Finally, based on the fusion of the material identification results, product identification results, and weight data, a determination is made as to whether the item is a qualified recyclable, and corresponding control commands are output. This effectively solves the technical problems of model coupling leading to update difficulties and serial identification failing to meet real-time requirements in existing recycling identification schemes. It achieves a technological breakthrough by simultaneously performing high-precision material identification and product identification on edge computing devices, significantly improving the accuracy and processing efficiency of fine-grained identification of recyclable items, while reducing model maintenance costs and hardware deployment barriers.

[0107] Figure 3 This application provides another schematic diagram of the process for identifying and processing recyclables, such as... Figure 3 As shown, in this embodiment... Figure 2 Based on the embodiments, the specific process of the recyclable identification and processing method is described in detail. The method includes:

[0108] S301: Continuously reads video frames from the camera through a background thread and updates the latest frame to the shared memory buffer in real time.

[0109] Specifically, the processor creates a separate background thread during system initialization. This thread continuously calls the camera's video capture interface in a loop to read the latest available video frame.

[0110] To prevent the default driver layer from using a multi-frame buffering mechanism that might cause the read frames to be outdated, the background thread immediately overwrites the original frame data in the shared memory buffer after each read, ensuring that the buffer always contains the most recently captured video frame.

[0111] The shared memory buffer serves as a medium for data exchange between threads. It is managed using a circular buffer or a single buffer with timestamps. The background thread is responsible for writing, and the main identification thread is responsible for reading.

[0112] This background frame washing mechanism decouples image acquisition from recognition processing, allowing the main recognition thread to avoid waiting for the frame synchronization delay of the camera hardware.

[0113] S302, In response to the trigger signal generated when the recycled item reaches the preset identification position, read the latest frame from the shared memory buffer as image data.

[0114] Specifically, the trigger signal is usually generated by an infrared sensor, photoelectric sensor or proximity sensor set in the recognition area. When an object blocks the sensor beam or triggers the sensing area, the sensor sends a level transition signal or bus message to the processor.

[0115] After receiving the trigger signal, the processor immediately reads the latest frame from the shared memory buffer. Since the background thread continuously refreshes the buffer, the frame read at this time is the image of the object arriving at the recognition location.

[0116] To ensure thread safety, read operations must be performed in conjunction with thread locking mechanisms, such as using mutex locks to protect buffer access, locking when a background thread writes, and locking when the main identification thread reads, to avoid data inconsistency caused by read-write conflicts.

[0117] The aforementioned asynchronous frame acquisition mechanism can solve the shutter delay and image transmission delay problems existing in the traditional trigger shooting mode, and is particularly suitable for real-time capture needs in continuous motion scenarios of conveyor belts.

[0118] Furthermore, the first recognition model is a pre-trained model for recognizing the material category of the items to be recycled, and the second recognition model is a pre-trained model for recognizing the product label category of the items to be recycled; the first recognition model and the second recognition model are stored as independent model files, and the second recognition model supports iterative updates independently of the first recognition model to adapt to newly added product label categories.

[0119] Specifically, the two models are stored as independent model files, using the Open Neural Network Exchange Format (ONE) to facilitate cross-platform deployment and version management.

[0120] When a new product identification category needs to be added, operators only need to retrain the second recognition model and replace the corresponding model file. The first recognition model remains completely unchanged, requiring neither retraining nor redeployment of the entire system. This model decoupling design avoids the material model retraining problem caused by brand expansion, significantly reducing model maintenance costs and iteration cycles.

[0121] Meanwhile, storing the two models independently also facilitates differentiated updates. For example, a second identification model containing different brand sets can be deployed for recycling programs in different regions, while the material model remains consistent.

[0122] In some embodiments, the processor converts the brand labels output by the model into communication protocol codes via a protocol mapping layer. The brand labels output by the model are in string format, and the protocol mapping layer converts them into corresponding hexadecimal codes, such as 0x05, according to a predefined mapping table, for communication with the hardware control module. This decoupled design ensures that algorithm updates do not affect the hardware control logic; adding a new brand only requires updating the mapping table without modifying the underlying control code.

[0123] Furthermore, the lightweight deployment of the first and second recognition models is achieved through the following steps:

[0124] Step A1: During the training phase, data augmentation strategies are used to simulate flattening, occlusion, and lighting changes in the recycling scene to train the first and second recognition models.

[0125] Specifically, data augmentation strategies include random rotation, random scaling, random cropping, color dithering, Gaussian noise addition, blurring simulation, and targeted augmentation such as flattening deformation simulation and local occlusion simulation.

[0126] For example, flattening and deformation simulation can be achieved through affine or perspective transformation of the image, compressing a normal can image vertically by a certain proportion to simulate recyclable items with different degrees of flattening. Occlusion simulation can be achieved by adding black rectangular blocks or interference areas cropped from other images at random locations in the image to simulate QR code occlusion or stain coverage. Illumination variation simulation can be achieved by adjusting image brightness, contrast, and saturation to simulate imaging effects under different time periods and ambient lighting conditions. Through these data augmentations, the trained model has a stronger generalization ability to various changes in real-world recycling scenarios.

[0127] In some embodiments, during training, the processor can employ a focus loss function to address the sample imbalance problem in material recognition. Since the number of samples for different material categories can vary significantly—for example, there are far more samples of pure aluminum cans than paper cans—using standard cross-entropy loss can lead to a model bias towards the majority class. Focus loss introduces a modulation factor on top of cross-entropy loss, reducing the weight of easily classified samples and causing the model to focus more on the minority class samples that are difficult to classify, thus improving the overall accuracy of material recognition.

[0128] In some embodiments, the processor can introduce an attention mechanism into the brand recognition model to enhance feature extraction of the logo region. The attention mechanism includes channel attention and spatial attention. Channel attention learns the weights of each feature channel to highlight the feature channels that contribute more to brand logo recognition. Spatial attention generates a two-dimensional attention mask, enabling the model to focus on the local region where the brand logo is located in the image, suppressing background interference and improving the accuracy and robustness of brand recognition.

[0129] In some embodiments, the processor may employ label smoothing and hybrid augmentation strategies to optimize model training. Label smoothing converts hard labels into soft labels, preventing the model from becoming overconfident and overfitting; hybrid augmentation mixes two training images in a certain proportion to generate new samples, and the corresponding labels are also mixed in the same proportion, increasing the diversity of training data and improving the model's generalization ability.

[0130] The above optimization strategies significantly improve the accuracy of material recognition validation sets and brand recognition validation sets.

[0131] Step A2: After training is complete, convert the trained floating-point model into an INT8 quantized model and calibrate it using a calibration set containing real-world scene samples until the accuracy loss of the quantized model is lower than the preset value.

[0132] Specifically, the quantization conversion uses post-training quantization technology, which means that there is no need to retrain, and the dynamic range of activation values ​​is directly used from the calibration set to determine the quantization parameters.

[0133] The calibration set should contain at least one thousand images collected from real-world recycling scenarios, covering samples with different lighting conditions, angles, and degrees of compression, to ensure that the quantized model can adapt to the actual deployment environment.

[0134] During the conversion process, the weights and activation values ​​in the floating-point model are mapped from 32-bit floating-point numbers to 8-bit integers, which can significantly reduce the model size and computational load, while accelerating inference speed.

[0135] The accuracy loss after quantization is usually controlled within 1%. By repeatedly adjusting the calibration set and quantization parameters, the best balance between accuracy and efficiency can be achieved.

[0136] In actual testing, after the above lightweighting process, the total number of parameters of the two models can be controlled within 3.5 million, the number of floating-point operations is less than 150 million, and the total size of the model file is compressed to less than 50MB, which meets the storage and memory limitations of embedded devices.

[0137] S303. Obtain the weight data of the items to be recycled.

[0138] Specifically, weight data is collected in real time by load cells, which typically use the resistance strain gauge principle and are installed below the conveyor mechanism. When an item is placed on the conveyor belt, the sensor outputs a voltage signal that is proportional to the weight, which is then amplified and converted into a digital value.

[0139] The sampling timing of weight data is synchronized with image acquisition, that is, weight reading is started immediately after the trigger signal is generated, to ensure that the weight data and image data correspond to the same item.

[0140] In some embodiments, after acquiring image data and weight data, the processor may also perform at least one of the following processing methods:

[0141] Case 1: Occlusion detection is performed based on image data, and a second control command is output when the occlusion rate of the key area of ​​the item to be recycled exceeds the preset occlusion threshold.

[0142] Specifically, occlusion detection can be achieved through an image segmentation model, which is specifically trained to identify key areas such as QR code areas and brand logo areas, and to calculate the proportion of occluded parts in these areas.

[0143] The occlusion rate is defined as the ratio of the number of occluded pixels to the total number of pixels in the key area. When the occlusion rate exceeds 40%, it is judged as malicious occlusion or unrecognizable, and a return command is directly output.

[0144] The occlusion detection model can employ a lightweight semantic segmentation network, which can run in parallel or sequentially with the main recognition model, and can be flexibly scheduled according to system resources.

[0145] Scenario 2: Based on image data, perform contour analysis, and when the flattening rate of the item to be recycled exceeds the preset flattening threshold, perform a downgrade operation on the item to be recycled.

[0146] Specifically, the contour analysis first extracts the outer contour of the product using an edge detection algorithm, then calculates the minimum bounding rectangle or ellipse of the contour, compares it with the aspect ratio of a standard can, and obtains the flattening ratio.

[0147] The flattening ratio is defined as (standard height - actual height) / standard height, or calculated using the area compression ratio.

[0148] When the flattening rate exceeds 70%, the item is still recyclable but has a lower value. The processor can classify it as a downgraded recyclable and perform downgrade processing operations such as halving the points or sorting it separately to a specific storage warehouse.

[0149] The specific rules for degradation processing can be configured according to the operation strategy, and the processor attaches a degradation flag when outputting control instructions.

[0150] Scenario 3: Match the image data with the pre-stored hazardous materials database and perform double verification based on abnormal weight data. If the match is successful and the weight exceeds the preset normal range, the item to be recycled will be identified as a hazardous material and an alarm or return operation will be triggered.

[0151] Specifically, the hazardous materials database contains image feature templates of common hazardous materials such as batteries, lighters, and power banks, and the matching algorithm can use feature point matching or deep feature comparison.

[0152] When visual matching is successful and the weight data deviates significantly from the normal weight range of recycled items, such as less than 5g or more than 500g, the processor determines it as a dangerous item, immediately triggers a return operation and records alarm information, and reports it to the management platform via network if necessary.

[0153] In some embodiments, the recycling machine is equipped with multiple infrared sensors that work together to monitor the item delivery process. Sensor 1 is located at the delivery port to detect whether an item has been inserted; sensor 2 is located at the visual recognition stop position to detect whether the item has reached the recognition location; and sensor 3 is located at the entrance of the collection compartment to detect whether the item has successfully fallen. If any sensor fails to receive a signal within the expected time after the user scans the code to deliver the item, the processor determines that the machine is malfunctioning or that there has been sabotage, sends an alarm message to the management platform, and suspends the service.

[0154] S304. Based on different threads in multi-threading technology, the inference tasks corresponding to the first recognition model and the second recognition model are run in the processor respectively.

[0155] Specifically, the processor creates two independent inference threads based on the number of cores, with each thread bound to a different CPU core to avoid cache invalidation caused by thread migration.

[0156] After the threads are created, they load the inference engine instances of the first and second recognition models respectively, and copy the preprocessed image data to their respective input buffers. Both threads start their inference tasks simultaneously, with each thread independently executing the model's forward propagation computation.

[0157] The inference engine employs a lightweight runtime framework optimized for the CPU instruction set and supports multi-threaded parallel execution.

[0158] S305. Based on the thread synchronization mechanism, after both the first and second recognition models have completed inference, output the material recognition result and the product identification result.

[0159] Specifically, thread synchronization can be achieved using barriers or condition variables. The main thread blocks by calling a wait function until both inference threads have completed their calculations and set a completion flag.

[0160] Once both threads have finished, the main thread reads the recognition results from the output buffers of the two models, including the category labels and the corresponding confidence scores.

[0161] Since the two models run in parallel, the total inference time is approximately equal to the inference time of the longer model in a single test. The actual test results show that it can be controlled within 220ms, which is significantly better than the sum of the time consumed by serial inference.

[0162] S306. Obtain the first confidence level of the material identification result and the second confidence level of the product identification result.

[0163] Specifically, the confidence score is the probability value of the model's output layer after normalization by the Softmax function. The value ranges from 0 to 1, and it represents the degree of certainty of the model regarding the current recognition result.

[0164] The first confidence level corresponds to the probability of predicting the material category, and the second confidence level corresponds to the probability of predicting the product identifier category. The processor reads these confidence level values ​​for subsequent fusion and judgment.

[0165] S307. Determine the corresponding preset weight range based on the material identification results and / or product identification results.

[0166] Specifically, the processor maintains a mapping table that records the standard weight range corresponding to different material categories and product labels. This mapping table can be pre-built based on historical statistical data and data provided by manufacturers, or it can be dynamically updated using machine learning methods. For example, the standard weight range for a brand's 500ml beverage can is 20g to 25g.

[0167] When both the material identification result and the product identification result are valid, the fine weight range corresponding to the product identification shall be used first; if the confidence level of the product identification is low or it cannot be identified, the coarse weight range corresponding to the material category shall be used instead.

[0168] S308. If the first confidence level is greater than the first preset threshold, the second confidence level is greater than the second preset threshold, and the weight data is within the preset weight range, then the item to be recycled is determined to be a qualified recyclable; otherwise, it is determined to be an unqualified recyclable.

[0169] Specifically, the first preset threshold can be set to 90%, and the second preset threshold can be set to 85%. The specific values ​​can be adjusted according to the actual application scenario. The weight range judgment adopts interval comparison, that is, the actual weight is greater than or equal to the lower limit and less than or equal to the upper limit.

[0170] Only items that meet both the confidence level requirement and the weight range requirement are considered qualified; items that fail to meet either requirement are considered unqualified.

[0171] Therefore, the multimodal fusion judgment mechanism effectively improves the recognition accuracy and reduces the risk of misjudgment of a single modality.

[0172] S309. If the product is determined to be a qualified recyclable, output the first control command; or if the product is determined to be an unqualified recyclable, output the second control command.

[0173] Specifically, the first control command instructs the conveyor mechanism to transport the items to be recycled to the collection bin and triggers reward feedback operations such as points distribution; the second control command instructs the conveyor mechanism to return the items to be recycled to the exit. The control commands are sent to the motor drive module via a general-purpose input / output interface or bus communication, driving the motor to rotate forward or reverse to achieve the sorting action.

[0174] The processor simultaneously records the identification results and processing actions, generating a transaction log for subsequent statistics and reconciliation.

[0175] The recyclable material identification and processing method provided in this disclosure solves the latency problem of fixed-point capture of moving items by ensuring the synchronization of trigger signals and image acquisition through an asynchronous frame acquisition mechanism; it solves the maintenance problem of retraining the entire system due to dynamic brand expansion by using a dual-model independent storage and iterative update design; it improves the ability to reject abnormal items through multimodal anomaly handling mechanisms such as occlusion detection, contour analysis, and hazardous material identification; it optimizes the efficiency of dual-model inference through parallel inference and thread synchronization mechanisms; and it improves the accuracy of qualified product determination by combining confidence threshold and dynamic weight range fusion judgment. Thus, it achieves full-process optimization from image acquisition, model deployment, anomaly detection, and parallel inference, ensuring the real-time performance, maintainability, and robustness of the recycling system in complex scenarios.

[0176] Figure 4 This is a schematic diagram of the structure of the recyclable identification and processing device provided in this application, as shown below. Figure 4 As shown, the recyclable material identification and processing device 400 provided in this embodiment is applied in the processor of a recycling machine and includes:

[0177] The acquisition module 410 is used to acquire image data and weight data of the items to be recycled;

[0178] The recognition module 420 is used to input image data into the first recognition model and the second recognition model respectively, perform inference in parallel, and output the corresponding material recognition result and product identification result; wherein, the material recognition result is output by the first recognition model and the product identification result is output by the second recognition model. The first recognition model and the second recognition model are models that are independently deployed in the processor after being lightweighted.

[0179] Analysis module 430 is used to determine whether the recycled items are qualified recyclables based on the material identification results, product identification results, and weight data;

[0180] The processing module 440 is used to output a first control command if the item is determined to be a qualified recyclable, or to output a second control command if the item is determined to be an unqualified recyclable; wherein, the first control command is used to instruct the conveying mechanism to convey the item to be recycled to the collection bin and trigger a reward feedback operation; the second control command is used to instruct the conveying mechanism to return the item to be recycled to the exit.

[0181] Optionally, the recognition module 420 is specifically used for: the first recognition model is a pre-trained model for recognizing the material category of the item to be recycled, and the second recognition model is a pre-trained model for recognizing the commodity label category of the item to be recycled; the first recognition model and the second recognition model are stored as independent model files, and the second recognition model supports iterative updates independently of the first recognition model to adapt to newly added commodity label categories.

[0182] Optionally, the recognition module 420 is specifically used to implement lightweight processing in the following way: during the training phase, a data augmentation strategy is used to simulate the flattening, occlusion and illumination changes in the recycling scene to train the first recognition model and the second recognition model; after training, the trained floating-point model is converted into an INT8 quantization model and calibrated using a calibration set containing real scene samples until the accuracy loss of the quantization model is lower than a preset value.

[0183] Optionally, the acquisition module 410 is specifically used to continuously read video frames from the camera through a background thread and update the latest frame to the shared memory buffer in real time; in response to the trigger signal generated when the recycled item reaches the preset recognition position, it reads the latest frame from the shared memory buffer as image data.

[0184] Optionally, the recognition module 420 is specifically used to run the inference tasks corresponding to the first recognition model and the second recognition model respectively in the processor based on different threads in multi-threading technology; and to output the material recognition result and the product identification result after the first recognition model and the second recognition model have completed inference based on the thread synchronization mechanism.

[0185] Optionally, the analysis module 430 is specifically used to: obtain a first confidence level of the material identification result and a second confidence level of the product identification result; determine the corresponding preset weight range based on the material identification result and / or the product identification result; if the first confidence level is greater than a first preset threshold, the second confidence level is greater than a second preset threshold, and the weight data is within the preset weight range, then the item to be recycled is determined to be a qualified recyclable; otherwise, it is determined to be an unqualified recyclable.

[0186] Optionally, the processing module 440 is further configured to, after acquiring the image data and weight data of the item to be recycled, perform at least one of the following actions: perform occlusion detection based on the image data, and output a second control command when the occlusion rate of the key area of ​​the item to be recycled exceeds a preset occlusion threshold; perform contour analysis based on the image data, and perform a downgrade operation on the item to be recycled when the flattening rate of the item to be recycled exceeds a preset flattening threshold; match the image data with a pre-stored hazardous materials database, and perform double verification in conjunction with abnormal weight data; if the match is successful and the weight exceeds a preset normal range, then the item to be recycled is identified as a hazardous material and an alarm or return operation is triggered.

[0187] The recyclable material identification and processing device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0188] Figure 5This is a schematic diagram of the structure of an electronic device provided in this application. Figure 5 As shown, the electronic device 50 includes:

[0189] Processor 51, memory 52, and communication interface 53;

[0190] Memory 52 is used to store executable instructions of processor 51;

[0191] The processor 51 is configured to execute the technical solutions in any of the foregoing method embodiments by executing the aforementioned executable instructions.

[0192] Optionally, the memory 52 can be either standalone or integrated with the processor 51.

[0193] Optionally, when the memory 52 is a device independent of the processor 51, the electronic device 50 may further include:

[0194] Bus 54, memory 52 and communication interface 53 are connected to processor 51 through bus 54 and complete communication with each other. Communication interface 53 is used to communicate with other devices.

[0195] Optionally, the communication interface 53 can be implemented using a transceiver. The communication interface is used to enable communication between the database access device and other devices (e.g., clients, read-write databases, and read-only databases). The memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk drive.

[0196] Bus 54 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0197] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0198] The electronic device is used to execute the technical solutions in any of the foregoing method embodiments. Its implementation principle and technical effect are similar, and will not be described again here.

[0199] This application also provides a readable storage medium storing a computer program thereon, which, when executed by a processor, implements the technical solutions provided in any of the foregoing method embodiments.

[0200] This application also provides a computer program product, including a computer program, which, when executed by a processor, is used to implement the technical solutions provided in any of the foregoing method embodiments.

[0201] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0202] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying and processing recyclables, characterized in that, The processor used in the recycling machine includes: Obtain image and weight data of the items to be recycled; The image data is input into the first recognition model and the second recognition model respectively, and inference is performed in parallel to output the corresponding material recognition result and product identification result; wherein, the material recognition result is output by the first recognition model and the product identification result is output by the second recognition model, and the first recognition model and the second recognition model are models that are deployed independently in the processor after being lightweighted; Based on the material identification results, the product identification results, and the weight data, it is determined whether the recycled item is a qualified recyclable item; If the item is determined to be a qualified recyclable, a first control command is output; or if the item is determined to be an unqualified recyclable, a second control command is output. The first control command instructs the conveying mechanism to transport the item to be recycled to the collection bin and triggers a reward feedback operation. The second control command instructs the conveying mechanism to return the item to be recycled to the exit.

2. The method according to claim 1, characterized in that, The first recognition model is a pre-trained model for recognizing the material category of the items to be recycled, and the second recognition model is a pre-trained model for recognizing the product label category of the items to be recycled. The first recognition model and the second recognition model are stored as independent model files, and the second recognition model supports iterative updates independently of the first recognition model to adapt to newly added product label categories.

3. The method according to claim 2, characterized in that, The lightweighting process is achieved in the following way: During the training phase, data augmentation strategies are used to simulate flattening, occlusion, and lighting changes in the recycling scenario to train the first and second recognition models. After training, the trained floating-point model is converted into an INT8 quantization model and calibrated using a calibration set containing real-world scene samples until the accuracy loss of the quantization model is lower than a preset value.

4. The method according to claim 1, characterized in that, The acquisition of image data of the items to be recycled includes: A background thread continuously reads video frames from the camera and updates the latest frame to the shared memory buffer in real time. In response to a trigger signal generated when the recycled item reaches a preset identification location, the latest frame is read from the shared memory buffer as the image data.

5. The method according to claim 1, characterized in that, The step of inputting the image data into the first recognition model and the second recognition model respectively, performing inference in parallel, and outputting the corresponding material recognition result and product identification result includes: Based on different threads in multithreading technology, the inference tasks corresponding to the first recognition model and the second recognition model are run in the processor respectively. Based on the thread synchronization mechanism, after both the first recognition model and the second recognition model have completed inference, the material recognition result and the product identification result are output.

6. The method according to claim 1, characterized in that, The step of determining whether the recycled item is a qualified recyclable based on the material identification result, the product identification result, and the weight data includes: Obtain the first confidence level of the material identification result and the second confidence level of the product identification result; Based on the material identification results and / or the product identification results, determine the corresponding preset weight range; If the first confidence level is greater than the first preset threshold, the second confidence level is greater than the second preset threshold, and the weight data is within the preset weight range, then the item to be recycled is determined to be a qualified recyclable; otherwise, it is determined to be an unqualified recyclable.

7. The method according to any one of claims 1 to 6, characterized in that, After obtaining the image data and weight data of the items to be recycled, the method further includes at least one of the following: Based on the image data, occlusion detection is performed, and when the occlusion rate of the key area of ​​the item to be recycled exceeds the preset occlusion threshold, the second control command is output. Based on the image data, contour analysis is performed, and when the flattening rate of the item to be recycled exceeds a preset flattening threshold, the item to be recycled is downgraded. The image data is matched with a pre-stored hazardous materials database, and double verification is performed based on abnormal weight data. If the match is successful and the weight exceeds the preset normal range, the item to be recycled is identified as a hazardous material and an alarm or return operation is triggered.

8. A recyclable material identification and processing device, characterized in that, The processor used in the recycling machine includes: The acquisition module is used to acquire image data and weight data of the items to be recycled; The recognition module is used to input the image data into the first recognition model and the second recognition model respectively, perform inference in parallel, and output the corresponding material recognition result and product identification result; wherein, the material recognition result is output by the first recognition model and the product identification result is output by the second recognition model, and the first recognition model and the second recognition model are models that have been lightweighted and deployed independently in the processor; The analysis module is used to determine whether the recycled item is a qualified recyclable based on the material identification result, the product identification result, and the weight data. The processing module is used to output a first control command if the item is determined to be a qualified recyclable, or to output a second control command if the item is determined to be an unqualified recyclable; wherein, the first control command is used to instruct the conveying mechanism to convey the item to be recycled to the collection bin and trigger a reward feedback operation; the second control command is used to instruct the conveying mechanism to return the item to be recycled to the exit.

9. An electronic device, characterized in that, include: Processor, memory, communication interface; The memory is used to store the executable instructions of the processor; The processor is configured to execute the recyclable identification processing method according to any one of claims 1 to 7 by executing the executable instructions.

10. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the recyclable identification and processing method according to any one of claims 1 to 7.

Images