An AI image recognition-based waste intelligent classification device and method

Abstract

The application discloses a kind of based on AI image recognition's waste intelligent classification equipment and method, belong to environmental protection equipment and artificial intelligence technical field;Equipment includes field classification device and cloud collaborative platform;Field classification device gathers image, extracts feature and carries out preliminary identification, if confidence is lower than first threshold value, then feature is uploaded;Cloud platform carries out depth identification, if its confidence is lower than second threshold value, then confirmation request containing reference information is provided to user;After user confirmation, this feedback and feature will be as training sample triggers the incremental learning of cloud model, and the updated model parameters are synchronized to field device again.The application realizes efficient, self-adapting, synergistically evolving waste intelligent classification by "field-cloud" double-layer collaborative decision and "man-in-loop" feedback evolution mechanism, significantly reduces the dependence on network, and continuously improves system recognition ability by minimizing user interaction.

Description

Technical Field

[0001] This invention belongs to the technical field of environmental protection equipment and artificial intelligence, and relates to an intelligent waste sorting device and method based on AI image recognition. Background Technology

[0002] With the implementation of waste sorting policies, intelligent waste sorting equipment has emerged. Most existing intelligent waste sorting equipment uses image recognition technology, and its implementation methods can be mainly divided into the following categories: First, relying on high-performance cloud servers for image recognition, and then returning the results to the device for sorting. This method requires high network stability, suffers from latency, and the uploading of all image data increases the computing and communication pressure on the cloud. Second, the device integrates offline recognition models, which reduces network dependence, but model updates are difficult and cannot adapt to changes in waste sorting rules and newly emerging waste types. Third, combining multiple sensors (such as weight and material sensors) for auxiliary judgment improves accuracy but increases hardware costs and complexity.

[0003] Furthermore, existing devices primarily use them for user interaction, such as authentication or requiring users to manually select a category when identification fails. The data value generated by this interaction is not being fully utilized to promote the system's own evolution. How to build an intelligent waste sorting system that can respond in real time, continuously learn and evolve, and efficiently utilize user feedback is a pressing issue that needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent waste sorting device and method based on AI image recognition. This device achieves efficient, adaptive, and evolvable intelligent waste sorting by constructing a two-layer decision-making architecture that coordinates on-site devices and a cloud platform, and by introducing a confidence-based intelligent routing and human-in-the-loop feedback evolution mechanism.

[0005] The technical solution adopted in this invention is as follows: In a first aspect, the present invention provides a waste intelligent sorting device based on AI image recognition, including an on-site sorting device and a cloud-based collaborative platform that is communicatively connected to the on-site sorting device; The on-site sorting device includes: The image acquisition module is configured to acquire images of the waste to be sorted. The preprocessing and feature extraction module is connected to the image acquisition module and is configured to preprocess the acquired images and run the first neural network model to extract the feature vectors of the images. The on-site decision-making module is connected to the preprocessing and feature extraction module and is configured to receive feature vectors and run a second neural network model to output the first classification result and the corresponding first confidence score based on the feature vectors. The actuator is configured to receive sorting instructions and place waste into the corresponding sorting bins; The first communication module connects to the on-site decision-making module and the execution mechanism; Cloud-based collaboration platforms include: The second communication module is connected to the first communication module. The knowledge base stores waste classification rules, a standard feature vector library, and related information. The model decision module is connected to the second communication module and the knowledge base. It is configured to run the third neural network model, perform deep recognition on the received feature vectors, and output the second classification result and the corresponding second confidence score. The confidence assessment and routing module is connected to the second communication module, the model decision module, and the knowledge base, respectively. The model evolution engine connects to the model decision module and the knowledge base; The feedback learning module connects to the model evolution engine, knowledge base, and second communication module. The on-site decision-making module is further configured to: compare a first confidence level with a preset on-site confidence level threshold; if the first confidence level is greater than or equal to the on-site confidence level threshold, generate a first control command containing the first classification result and send it to the execution mechanism through the first communication module; if the first confidence level is less than the on-site confidence level threshold, generate a cloud assistance request containing feature vectors and send it to the cloud collaboration platform through the first communication module. The confidence assessment and routing module is configured as follows: it receives the feature vector from the cloud assistance request and sends it to the model decision module; it receives the second classification result and second confidence level output by the model decision module; it compares the second confidence level with a preset cloud confidence threshold; if the second confidence level is greater than or equal to the cloud confidence threshold, it generates a second control command containing the second classification result and sends it to the actuator through the second communication module and the first communication module; if the second confidence level is less than the cloud confidence threshold, it generates a user confirmation request, which includes the second classification result and reference information related to the second classification result obtained from the knowledge base, and sends it to the on-site classification device for prompting through the second communication module and the first communication module. The on-site sorting device also includes a human-computer interaction module, which is connected to the first communication module and configured to receive and display user confirmation requests, and to receive user confirmation instructions input by the user based on the user confirmation requests; The feedback learning module is configured to: receive user confirmation instructions and corresponding feature vectors through the second communication module and the first communication module; use the user confirmation instructions as the annotation results of the feature vectors, form training sample pairs with the feature vectors, store them in the knowledge base, and trigger the model evolution engine; The model evolution engine is configured to: respond to the triggering of the feedback learning module, use the newly added training sample pairs in the knowledge base to perform incremental learning training on the third neural network model, and update the parameters of the third neural network model; The second communication module is also configured to synchronize some or all of the updated parameters of the third neural network model to the first communication module of one or more field classification devices to update the corresponding second neural network model.

[0006] Furthermore, preprocessing includes at least one operation on the image, such as size normalization, grayscale conversion, or noise filtering.

[0007] Furthermore, the on-site confidence threshold is higher than the cloud-based confidence threshold.

[0008] Furthermore, the user confirmation request includes reference information such as example images and / or text descriptions of other wastes belonging to the same category as the second classification result.

[0009] Furthermore, the on-site sorting device also includes a waste carrying and transfer mechanism. An image acquisition module is set above or to the side of the waste carrying and transfer mechanism. An actuator is connected to the waste carrying and transfer mechanism and configured to drive the waste carrying and transfer mechanism to move so as to transfer the waste to the top of the target sorting bin and place it.

[0010] Furthermore, the human-computer interaction module can be a touch screen, a voice interaction unit, or an application programming interface (API) bound to a user's mobile terminal.

[0011] Secondly, the present invention provides a waste intelligent classification method based on AI image recognition, applied to the device described above, the method comprising: S1. Acquire images of the waste to be sorted using the image acquisition module of the on-site sorting device; S2. The image is preprocessed using the preprocessing and feature extraction module, and the feature vector of the image is extracted. S3. The first classification result and the first confidence level are obtained based on the feature vector through the on-site decision-making module; S4. Determine whether the first confidence level is greater than or equal to the on-site confidence threshold; If so, then execute S5: The on-site decision-making module generates the first control command, which controls the actuator to dispose of the waste according to the first classification result; If not, then execute S6: The on-site decision-making module uploads the feature vector to the cloud collaboration platform; S7. The model decision module of the cloud-based collaborative platform performs deep recognition based on the received feature vectors to obtain the second classification result and the second confidence level; S8. Determine whether the second confidence level is greater than or equal to the cloud confidence threshold; If so, then execute S9: The cloud-based collaborative platform generates a second control command and sends it to the on-site sorting device to control the actuator to dispose of the waste according to the second sorting result; If not, then execute S10: The cloud collaboration platform generates a user confirmation request and sends it to the human-machine interaction module of the on-site classification device for prompting. The user confirmation request includes the second classification result and related reference information. S11. Receive user confirmation commands input by the user through the human-computer interaction module; S12. Upload the user confirmation command and the corresponding feature vector as training sample pairs to the cloud collaboration platform to trigger the model evolution engine to perform incremental learning training on the third neural network model; S13. The cloud-based collaborative platform synchronizes the updated model parameters to one or more field classification devices to update their second neural network models.

[0012] Furthermore, in step S10, if the user does not input a confirmation command through the human-machine interaction module within a preset time, the control actuator will dispose of the waste into the unsorted bin or the temporary storage bin.

[0013] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described above.

[0014] The working principle of this invention lies in constructing a collaborative intelligent closed loop of "perception-decision-feedback-evolution", and the specific workflow is as follows: 1. Local Perception and Preliminary Decision-Making: When a user throws in waste, the on-site sorting device is immediately activated; the image acquisition module captures an image of the waste; the preprocessing and feature extraction module standardizes the image and uses a lightweight first neural network model (such as MobileNet) to extract a feature vector that represents the visual essence of the waste; subsequently, the on-site decision-making module uses a smaller second neural network model to quickly analyze the feature vector, obtaining a preliminary classification result (first classification result) and the degree of confidence in the result (first confidence level).

[0015] 2. Confidence-Based Intelligent Routing: The system introduces key routing logic. The on-site decision-making module compares a first confidence level with a higher "on-site confidence threshold." If the confidence level is high enough, indicating that the local model is very confident, control commands are directly generated to drive the actuators to complete the sorting and disposal process. The entire process is fast and does not rely on the network. If the confidence level is below the threshold, it indicates that the waste's characteristics are ambiguous or belong to the knowledge blind spot of the local model. In this case, no final decision is made, and only the extracted "feature vector"—the core data (instead of the original image, reducing the data volume)—is uploaded to the cloud collaboration platform through the first communication module to request assistance. This effectively filters out simple samples and concentrates computing and communication resources on difficult samples.

[0016] 3. Cloud-based deep identification and adjudication: After receiving the feature vector, the model decision module of the cloud-based collaborative platform calls its more powerful and complex third neural network model (such as ResNet, VisionTransformer, etc.) to perform deep analysis and adjudication, and outputs a more authoritative cloud classification result (second classification result) and the corresponding confidence level (second confidence level).

[0017] 4. Human-in-the-loop final decision and data collection: The confidence assessment and routing module compares the second confidence level in the cloud with a lower "cloud confidence threshold." If the confidence level meets the threshold, the cloud directly issues a classification instruction. If the confidence level of the cloud model is also insufficient (i.e., neither the field model nor the cloud model can be certain), the system does not abandon or process randomly, but instead activates the "human-in-the-loop" mechanism. The cloud platform retrieves reference information (such as images and text descriptions) related to the category it deems most likely from the knowledge base and generates a user confirmation request, which is then sent to the human-machine interface of the field device. By viewing the request and reference information, the user can quickly make a correct judgment (confirm or correct). This interaction step is extremely brief but generates crucial value: every click by the user provides accurate human annotation for a "difficult" feature vector.

[0018] 5. Feedback-driven model evolution: The user confirmation command and the previously uploaded feature vector form a high-quality "feature-label" training sample pair in the cloud feedback learning module. This sample pair is stored in the knowledge base and immediately (or in batches) triggers the model evolution engine. The engine uses this new sample to perform rapid incremental learning (fine-tuning) on ​​the cloud-based third neural network model, enabling the model to remember this new feature and thus improve the recognition confidence when encountering similar features in the future. This gives the system the ability to continuously learn and adaptively optimize from real-world usage scenarios.

[0019] 6. Collaborative Synchronization of Swarm Intelligence: After the cloud-based model evolves, its learning results (updated model parameters) are dynamically synchronized to one, several, or all of the field classification devices in the network through the second communication module, updating their local second neural network models. This means that user feedback and model improvements obtained by any device can benefit the entire device network, achieving collaborative evolution of swarm intelligence. The next time similar waste is encountered, the local recognition capabilities of all devices have been improved, and they may be able to complete the classification locally with high confidence, forming a positive cycle of "the more it is used, the smarter it becomes".

[0020] The entire working principle tightly couples distributed field devices with a centralized cloud brain and user knowledge through two-layer decision filtering, confidence routing, collection of high-quality data with minimal human-computer interaction, real-time feedback learning and dynamic synchronization, forming a highly efficient, adaptive and evolvable complete intelligent system.

[0021] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. A waste intelligent sorting device and method based on AI image recognition, which sets up a lightweight recognition model at the field and performs confidence judgment, and only low-confidence samples are uploaded to the cloud for in-depth processing; this effectively reduces the amount of network data transmission and cloud computing load, reduces the system's absolute dependence on network real-time performance, and improves the overall response efficiency and robustness.

[0022] 2. In this invention, the user's "quick confirmation" behavior when faced with low-confidence recognition results is transformed into a source of high-quality training data. Through the "feedback learning module" and the "model evolution engine", the system can use this data to continuously perform online incremental learning, enabling the recognition model to quickly adapt to new waste types or changing classification rules, and thus possessing self-evolution capabilities.

[0023] 3. In this invention, after the cloud model evolves, the evolution results (model parameters) can be dynamically synchronized to all networked field devices; this means that user feedback and experience obtained by any device can benefit the entire device network, achieving a collaborative intelligent effect of "learning in one place and upgrading the whole network", which greatly improves the overall maintenance efficiency and intelligence level of the system.

[0024] 4. In this invention, when user intervention is required, relevant reference information (such as pictures and text descriptions) is provided to assist users in making accurate judgments, reducing the difficulty of user decision-making; at the same time, each user confirmation directly contributes to the improvement of the system, forming a positive incentive, enhancing user participation and the long-term reliability of the system. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments, experimental examples, and comparative examples will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 , Figure 2 This is a schematic diagram of the architecture of the device of the present invention; Figure 3 This is a schematic flowchart of the method of the present invention; Figure 4 This is a schematic diagram of the architecture of the preprocessing module in the device of the present invention; Figure 5 This is a schematic diagram of the architecture of the on-site sorting device in the equipment of the present invention; Figure 6 This is a schematic diagram of the human-computer interaction module in the device of the present invention; Figure 7 This is a schematic diagram of the architecture of step S10 in the method of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings, embodiments, experimental examples, and comparative examples. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described embodiments are merely some embodiments of the invention, and not all embodiments. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0027] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0028] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0029] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0030] I. Implementation Examples Example 1 This invention discloses an intelligent waste sorting device and method based on AI image recognition, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the specific implementation of this embodiment includes multiple on-site sorting devices deployed within the community, and a centralized cloud-based collaborative platform. The on-site sorting devices and the cloud-based collaborative platform communicate via the Internet.

[0031] The on-site sorting device includes a housing, a control motherboard installed inside the housing, an image acquisition module, a preprocessing and feature extraction module, an on-site decision-making module, a first communication module, a human-machine interaction module, an actuator, and a waste carrying and transfer mechanism.

[0032] The image acquisition module uses a high-definition RGB camera, which is fixedly installed above the carrying platform of the waste carrying and transfer mechanism to capture images of the top of the waste placed on the carrying platform.

[0033] The preprocessing and feature extraction module, along with the on-site decision-making module, are integrated into the processor of the control motherboard. The preprocessing and feature extraction module runs a lightweight convolutional neural network, which is trained to extract a fixed-dimensional feature vector from the input RGB image, rather than directly classifying it. Before extraction, the image is size normalized and simple illumination compensation is performed.

[0034] The on-site decision-making module runs a shallow fully connected neural network. This model takes a feature vector as input and outputs a vector whose dimension equals the number of waste classification categories. Each dimension's value represents the probability of belonging to that category. The category with the highest probability is taken as the first classification result, and this highest probability value is used as the first confidence level. The preset on-site confidence threshold within the on-site decision-making module is 0.85.

[0035] The waste carrying and transfer mechanism includes a translational slide driven by a stepper motor, and an openable and closable carrying hopper on the slide. The actuator includes a motor driver for driving the translational slide and a solenoid valve for controlling the opening and closing of the bottom hatch of the carrying hopper. The human-machine interface module is an embedded touchscreen mounted on the front of the housing. The first communication module is a wireless communication module.

[0036] The cloud-based collaborative platform is deployed on a cloud server and includes a second communication module, a knowledge base, a model decision-making module, a confidence assessment and routing module, a model evolution engine, and a feedback learning module.

[0037] The knowledge base is built using relational and vector databases. It stores a standard waste classification directory, text descriptions of common items in each category, example images, and their corresponding standard feature vectors. It also stores historically collected training sample pairs.

[0038] The model decision module runs a deep neural network with more complex parameters, achieving higher recognition accuracy than the second neural network model at the field site. Its input is also the feature vector extracted from the field site, and it outputs a second classification result and a second confidence score. The cloud-based confidence threshold is preset to 0.70.

[0039] The confidence assessment and routing module handles the logical control. The model evolution engine employs an online learning algorithm, enabling rapid fine-tuning of the deep neural network using new samples. The feedback learning module processes user feedback data.

[0040] Example 2 This invention discloses an intelligent waste sorting device and method based on AI image recognition, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the specific implementation method of this embodiment is as follows: The main difference between this embodiment and embodiment 1 lies in the variation of user interaction and processing logic.

[0041] In the steps of Example 1, if the user sees a prompt on the touch screen of the human-computer interaction module and believes that the system has made a mistake in recognition, the user can click the "Correct" button and select the correct category from the pop-up category list.

[0042] At this point, the training sample pairs received by the feedback learning module will be feature vectors and user-corrected category labels. The model evolution engine will use these corrected samples for training, thereby correcting the model's misperceptions of such easily confused items. This greatly accelerates the model's learning process for "long-tail" or special items.

[0043] Example 3 This invention discloses an intelligent waste sorting device and method based on AI image recognition, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the specific implementation of this embodiment is as follows: the difference between this embodiment and embodiment 1 lies in the structural variation of the field device.

[0044] The waste carrying and transfer mechanism can be a rotating pallet with multiple sorting bins arranged in a ring around it. An actuator drives the pallet to rotate, aligning the carrying hopper with the target sorting bin and then tilting it. An image acquisition module can be mounted on the side, working in conjunction with a linear scanning mechanism to scan the waste from multiple angles after it is deposited, acquiring multiple images. A preprocessing and feature extraction module can integrate features from multiple images to generate a more comprehensive fused feature vector, which is then fed into the on-site decision-making module. This helps improve the initial identification capability of waste with non-standard postures.

[0045] Example 4 This invention discloses an intelligent waste sorting device and method based on AI image recognition, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the specific implementation method of this embodiment is as follows: the difference between this embodiment and embodiment 1 lies in the variant of confidence assessment.

[0046] The on-site confidence threshold and the cloud-based confidence threshold do not need to be fixed values, but rather dynamically adjusted. For example, the cloud-based collaborative platform can statistically analyze the average network latency of different on-site devices at different times in historical data. When a high network latency is detected for a particular on-site device, a command can be sent via a second communication module to temporarily lower its on-site confidence threshold, encouraging more samples to be processed locally to mitigate the impact of network latency and ensure a better user experience. Once the network conditions are restored, the threshold is adjusted back to its original value.

[0047] Example 5 This invention discloses an intelligent waste sorting device and method based on AI image recognition, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 As shown, the specific implementation method of this embodiment is as follows: The difference between this embodiment and embodiment 1 lies in the variant of the model synchronization content.

[0048] Considering the limited storage and computing power of field devices, the cloud-based collaborative platform may not synchronize all model parameters during updates. After training, the model evolution engine can analyze the parameters that have changed most significantly due to the incremental learning. During synchronization, only this subset of "significantly changed" parameters, along with their location identifiers within the network, is sent to the field device. Upon receiving this information, the field device only updates the parameters at the corresponding locations in its second neural network model. This approach further reduces the amount of data synchronized and improves efficiency.

[0049] II. Experimental Examples and Comparative Examples Experiment Example 1: Verification of Core Process and Evolution Effect This experimental example simulates a specific scenario to verify the effectiveness of the complete closed loop of "local-cloud-feedback-evolution-synchronization" of the present invention.

[0050] Experimental setup: One on-site classification device of this invention (equipped with a lightweight MobileNetV2 for feature extraction and a shallow fully connected network for local decision-making) and a cloud platform (equipped with ResNet50 for deep decision-making) were deployed. The initial model was trained on a standard garbage dataset. The on-site confidence threshold was set to 0.85, and the cloud confidence threshold was set to 0.70. A severely soiled and flattened "lactic acid bacteria beverage bottle" (belonging to recyclable materials) that the model had not encountered before was introduced.

[0051] Experimental procedure: 1. Initial Deployment: The device collects images, and the local model extracts features, classifying the material as "recyclable," but with a confidence level of only 0.78 (<0.85). The feature vectors are uploaded to the cloud. After analysis by ResNet50 in the cloud, a "recyclable" classification is provided with a confidence level of 0.65 (<0.70). A user confirmation request (including a description of "recyclable" and an image of a clean beverage bottle) is generated and sent from the cloud. The user reviews and confirms that the information is correct. Deployment complete.

[0052] 2. Feedback and Evolution: The "feature vector-recyclable material" sample pair of this interaction triggers the cloud-based model evolution engine to perform an incremental learning of ResNet50.

[0053] 3. Model synchronization: The evolved model parameters (mainly the fully connected layers) are synchronized to the field device.

[0054] 4. Second placement: Place a second soiled beverage bottle of the same type. The updated local model directly provides a "recyclable" classification with a confidence level of 0.91 (>0.85), and the placement is completed locally without cloud assistance.

[0055] Experimental results: For items that are initially difficult to identify, this invention improves the system's confidence in identifying such features from 0.78 to 0.91 through a single user interaction, realizing the transformation from "requiring cloud assistance" to "local rapid decision-making", and proving the effectiveness of the feedback evolution closed loop.

[0056] Experiment Example 2: Comparison Experiment of System Efficiency and Resource Consumption This experimental example compares the efficiency of the present invention's solution with that of the "fully cloud-based recognition solution" in typical use cases.

[0057] Experimental setup: The solution of this invention is as set up as in Experiment Example 1.

[0058] Comparative Example 1 (Completely Cloud-Based Recognition): The on-site device is only responsible for taking photos and uploading the original images (compressed to 200KB / image); all recognition is completed in the cloud. The same cloud-based ResNet50 model is used.

[0059] Test dataset: 1000 junk images selected from a public dataset, including 80% common items (which can be identified with high confidence locally) and 20% suspicious items (which require user interaction).

[0060] Evaluation metrics: average total time for a single recognition (from taking a picture to issuing a command), average amount of data uploaded per instance, and cloud server computing load.

[0061] Table 1 shows the experimental results of Experiment Example 2. Results Analysis: This invention's solution processes 80% of high-confidence samples locally, avoiding the uploading of original images and cloud computing, resulting in a 62% reduction in average processing time, a 92% reduction in network traffic, and an 80% reduction in cloud computing load. This significantly improves system response speed and reduces network dependence and operating costs.

[0062] Experiment Example 3: Evolution Experiment of Long-Tail Item Recognition Ability This experimental example verifies the system's ability to continuously learn about uncommon (long-tail) waste categories.

[0063] Experimental setup: Over a one-month period, a mixed waste stream containing 5% newly appearing or extremely rare items (such as packaging of a certain new composite material or specific electronic waste) was continuously fed into the system of this invention (networked with 10 devices) and the comparative example 2 (two-layer recognition without feedback evolution) system (same initial model, with two-layer routing but no learning triggered after user confirmation).

[0064] Evaluation metric: Weekly statistics on the overall recognition accuracy of the system for the 5% of long-tail items (the proportion of items correctly classified without user correction).

[0065] Table 2 shows the experimental results of Experiment Example 3. Results Analysis: Through continuous learning from user feedback, the accuracy of long-tail item identification in this invention increased from 68% to 92% within four weeks. In contrast, the accuracy of Comparative Example 2 stagnated. This demonstrates that the "human-in-the-loop" feedback evolution mechanism is crucial for enabling the system to adapt to the ever-changing composition of waste in the real world—an "unexpected" effect that cannot be naturally obtained by those skilled in the art simply by setting up a two-layer architecture.

[0066] Experiment Example 4: Verification of the Collaborative Effect of Swarm Intelligence This experimental example verifies how synchronized model updates improve the overall performance of the device network.

[0067] Experimental Setup: Three field devices of this invention, A, B, and C, are connected to the network. First, a specific, difficult-to-identify item (such as an oil-stained paper lunchbox) is repeatedly introduced into device A, triggering multiple user confirmations and cloud-based learning. After learning is complete, the cloud updates the parameters and synchronizes them to devices B and C. Then, the local recognition confidence levels of devices A, B, and C for this difficult-to-identify item are tested respectively. Simultaneously, Comparative Example 3 (feedback learning without model synchronization) is set up, where after device A learns, the knowledge is not synchronized to devices B and C.

[0068] Evaluation metric: The confidence level of local recognition of the item by devices A, B, and C after learning synchronization.

[0069] Table 3 shows the experimental results of Experiment Example 4. Results Analysis: In the present invention, devices B and C acquired recognition capabilities (confidence > 0.85) through parameter synchronization without touching the item. In contrast, in Comparative Example 3, devices B and C maintained their initial low confidence levels. This demonstrates the synergistic effect of "learning in one place, upgrading the entire network," which greatly improves the operational efficiency and intelligence level of the entire device network—an effect that cannot be achieved by simply combining multiple independent intelligent devices.

[0070] Experiment Example 5: System Robustness Test (Weak Network Environment) This experimental example verifies the performance of the invention under poor network conditions.

[0071] Experimental setup: Simulate an environment where network latency increases from 100ms to 2000ms. Compare the processing success rate and average waiting time of the proposed solution with Comparative Example 1 (fully cloud-based recognition). Define "successful processing" as obtaining a classification result within 5 seconds acceptable to the user (whether it is a local decision, a cloud decision, or a decision made after user confirmation).

[0072] Evaluation metrics: Success rate and average processing time under different network latencies.

[0073] Experimental results: When the network latency is <500ms, the success rate of both schemes is close to 100%, and the average time consumption of the present invention is better. When network latency rises to 1500ms, compared to Example 1, where all requests rely on the cloud, the success rate drops to about 65%, and the average time exceeds 3 seconds. In contrast, the solution of this invention maintains a success rate of over 95% because more than 80% of requests are processed directly locally with high confidence, and the average time is less affected by the network.

[0074] Results Analysis: This invention digests most requests locally through confidence-based routing, significantly reducing the absolute dependence on network real-time performance and improving the robustness and availability of the system in complex network environments. This is an advantage that a fully cloud-based architecture cannot match.

[0075] Summary of scale settings: Comparative Example 1: A fully cloud-based identification scheme; used to highlight the advantages of this invention in terms of efficiency, energy consumption, network dependence, and robustness.

[0076] Comparative Example 2: A two-layer recognition scheme with no feedback evolution; used to highlight the key role of the "human-in-the-loop" feedback evolution mechanism of the present invention in realizing the system's adaptability and continuous learning ability, and to prove that this combination produces new effects that surpass the simple two-layer architecture.

[0077] Comparative Example 3: Feedback learning scheme without model synchronization; used to highlight the value of swarm intelligence collaborative evolution achieved by the "model parameter synchronization" mechanism of this invention, and to prove that it produces a synergistic effect of improving the overall network efficiency.

[0078] Comparative Example 4: Fixed Threshold Routing Scheme. That is, the confidence threshold is a fixed value. In contrast, the threshold in Example 4 of this invention can be dynamically adjusted. This illustrates the possibility of further optimization of this invention, namely, dynamically adjusting the routing strategy according to the network state, thereby maintaining superior performance in changing environments and demonstrating the system's intelligence level.

[0079] Conclusion: The above experimental examples and comparative examples show that the present invention introduces features such as "confidence-based intelligent routing", "human-in-the-loop feedback evolution" and "parameter synchronization of swarm intelligence", and forms a tightly coupled and mutually reinforcing new relationship under the "on-site-cloud" collaborative framework, thereby producing a synergistic effect of "1+1>2" in multiple dimensions such as system efficiency, adaptive learning ability, swarm intelligence enhancement and operational robustness.

[0080] The above description is only a preferred embodiment, experimental example, and comparative example of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A waste intelligent sorting device based on AI image recognition, characterized in that, Includes an on-site sorting device and a cloud-based collaborative platform that communicates with the on-site sorting device; The on-site sorting device includes: The image acquisition module is configured to acquire images of the waste to be sorted. The preprocessing and feature extraction module is connected to the image acquisition module and is configured to preprocess the acquired image and run a first neural network model to extract the feature vector of the image. The on-site decision-making module is connected to the preprocessing and feature extraction module and is configured to receive the feature vector and run a second neural network model to output a first classification result and a corresponding first confidence level based on the feature vector. The actuator is configured to receive sorting instructions and place waste into the corresponding sorting bins; The first communication module is connected to the on-site decision-making module and the execution mechanism; The cloud-based collaboration platform includes: The second communication module is communicatively connected to the first communication module. The knowledge base stores waste classification rules, a standard feature vector library, and related information. The model decision module is connected to the second communication module and the knowledge base, and is configured to run a third neural network model to perform deep recognition on the received feature vectors and output a second classification result and the corresponding second confidence level. The confidence assessment and routing module is connected to the second communication module, the model decision module, and the knowledge base, respectively. A model evolution engine, connected to the model decision module and the knowledge base; The feedback learning module is connected to the model evolution engine, the knowledge base, and the second communication module. The on-site decision-making module is further configured to: compare the first confidence level with a preset on-site confidence level threshold; if the first confidence level is greater than or equal to the on-site confidence level threshold, generate a first control command containing the first classification result and send it to the execution mechanism through the first communication module; if the first confidence level is less than the on-site confidence level threshold, generate a cloud assistance request containing the feature vector and send it to the cloud collaboration platform through the first communication module. The confidence assessment and routing module is configured to: receive the feature vector from the cloud assistance request and send it to the model decision module; receive the second classification result and second confidence level output by the model decision module; compare the second confidence level with a preset cloud confidence threshold; if the second confidence level is greater than or equal to the cloud confidence threshold, generate a second control command containing the second classification result and send it to the execution mechanism through the second communication module and the first communication module; if the second confidence level is less than the cloud confidence threshold, generate a user confirmation request, which includes the second classification result and reference information related to the second classification result obtained from the knowledge base, and send it to the on-site classification device for prompting through the second communication module and the first communication module. The on-site classification device further includes a human-computer interaction module connected to the first communication module, configured to receive and display the user confirmation request, and receive user confirmation instructions input by the user based on the user confirmation request; The feedback learning module is configured to: receive the user confirmation instruction and the corresponding feature vector through the second communication module and the first communication module; use the user confirmation instruction as the annotation result of the feature vector, form a training sample pair with the feature vector, store it in the knowledge base, and trigger the model evolution engine; The model evolution engine is configured to: respond to the triggering of the feedback learning module, use the newly added training sample pairs in the knowledge base to perform incremental learning training on the third neural network model, and update the parameters of the third neural network model; The second communication module is further configured to synchronize some or all of the updated parameters of the third neural network model to the first communication module of one or more of the field classification devices to update the corresponding second neural network model.

2. The intelligent waste sorting device based on AI image recognition according to claim 1, characterized in that, The preprocessing includes at least one operation of image size normalization, grayscale conversion, or noise filtering.

3. The intelligent waste sorting device based on AI image recognition according to claim 1, characterized in that, The on-site confidence threshold is higher than the cloud-based confidence threshold.

4. The intelligent waste sorting device based on AI image recognition according to claim 1, characterized in that, In the user confirmation request, the reference information includes example images and / or text descriptions of other wastes belonging to the same category as the second classification result.

5. The intelligent waste sorting device based on AI image recognition according to claim 1, characterized in that, The on-site sorting device also includes a waste carrying and transfer mechanism. The image acquisition module is located above or to the side of the waste carrying and transfer mechanism. The actuator is connected to the waste carrying and transfer mechanism and is configured to drive the waste carrying and transfer mechanism to move so as to transfer the waste to the top of the target sorting bin and place it.

6. The intelligent waste sorting device based on AI image recognition according to claim 1, characterized in that, The human-computer interaction module is a touch screen, a voice interaction unit, or an application programming interface (API) bound to a user's mobile terminal.

7. A waste intelligent classification method based on AI image recognition, applied to the equipment as described in any one of claims 1 to 6, characterized in that, The method includes: S1. Acquire images of the waste to be sorted using the image acquisition module of the on-site sorting device; S2. The image is preprocessed using the preprocessing and feature extraction module, and the feature vector of the image is extracted. S3. The first classification result and the first confidence level are obtained based on the feature vector through the on-site decision-making module; S4. Determine whether the first confidence level is greater than or equal to the on-site confidence threshold; If so, then execute S5: The on-site decision-making module generates the first control command, which controls the actuator to dispose of the waste according to the first classification result; If not, then execute S6: The on-site decision-making module uploads the feature vector to the cloud collaboration platform; S7. The model decision module of the cloud-based collaborative platform performs deep recognition based on the received feature vectors to obtain the second classification result and the second confidence level; S8. Determine whether the second confidence level is greater than or equal to the cloud confidence threshold; If so, then execute S9: The cloud-based collaborative platform generates a second control command and sends it to the on-site sorting device to control the actuator to dispose of the waste according to the second sorting result; If not, then execute S10: The cloud collaboration platform generates a user confirmation request and sends it to the human-computer interaction module of the on-site classification device for prompting. The user confirmation request includes the second classification result and related reference information. S11. Receive user confirmation commands input by the user through the human-computer interaction module; S12. Upload the user confirmation command and the corresponding feature vector as training sample pairs to the cloud collaboration platform to trigger the model evolution engine to perform incremental learning training on the third neural network model; S13. The cloud-based collaborative platform synchronizes the updated model parameters to one or more field classification devices to update their second neural network models.

8. The intelligent waste sorting method based on AI image recognition according to claim 7, characterized in that, In step S10, if the user does not input a confirmation command through the human-machine interaction module within a preset time, the control actuator will dispose of the waste into the unsorted bin or the temporary storage bin.

Images