A computer vision-based household garbage disposal type recommendation method and system

Abstract

The application discloses a computer vision-based household garbage throwing type recommendation method and system, which comprises the following steps: collecting multi-source data in the pre-divided area around the garbage recycling bin in real time; detecting and tracking the personnel approaching the garbage recycling bin based on the collected data, performing personnel action intention judgment and prediction; when the personnel approaching the garbage recycling bin enters a preset partition or the personnel action judgment and prediction shows a throwing intention, triggering recognition and triggering different recognition strategies according to different confidence levels of entering different preset partitions or judging and predicting the throwing intention; recognizing the hand area of the personnel, and segmenting the handheld object; fusing the multi-modal data collected synchronously to recognize the garbage type of the handheld object; and mapping the garbage type to a preset garbage classification standard to generate a throwing type recommendation result of the handheld object. The application can provide timely and accurate throwing type recommendations, improve the accuracy and efficiency of garbage classification, and improve the user experience.

Description

Technical Field

[0001] This application relates to the technical field of waste treatment and intelligent identification, specifically to a method and system for recommending types of household waste disposal based on computer vision. Background Technology

[0002] With increasing environmental awareness, waste sorting has become a focus of social attention. Effective waste sorting helps with resource recycling, reduces environmental pollution, and is of great significance to sustainable development. The emergence of intelligent waste sorting systems has, to some extent, promoted the implementation of waste sorting work and provided technical support for the efficiency and precision of waste treatment.

[0003] Traditional waste sorting relies primarily on basic methods. On one hand, static pictorial explanations are common, with classification standards and example pictures posted next to trash cans, allowing residents to determine the correct waste type. On the other hand, passive query systems are also widely used, allowing residents to check waste classification information by scanning codes or entering keywords. Furthermore, while existing smart trash cans are equipped with some sensing capabilities, they mainly focus on post-disposal identification and sorting, offering limited assistance and guidance before waste disposal.

[0004] However, these existing technologies have significant drawbacks. They lack initiative and real-time capability, and cannot provide targeted guidance based on the actual amount of waste residents possess. When disposing of waste, especially when holding multiple different types, residents still find it difficult to quickly determine the correct disposal port, leading to frequent mis-disposal and reducing the efficiency and accuracy of waste sorting. Summary of the Invention

[0005] To provide timely and accurate recommendations for waste disposal types, improve the accuracy and efficiency of waste sorting, and enhance user experience, this application provides a computer vision-based method and system for recommending waste disposal types.

[0006] Firstly, this application provides a computer vision-based method for recommending types of household waste disposal, including: Multi-source data is collected in real time within a pre-defined area around the waste recycling bin using multi-modal sensors; Based on the collected data, the system detects and tracks people approaching the recycling bins, and performs judgment and prediction of people's action intentions. When people approaching the recycling bins enter a preset zone or when the judgment and prediction of people's action intentions are made, the system triggers identification and triggers different identification strategies based on the different preset zones entered or the different confidence levels of the judgment and prediction of the disposal intentions. The identification strategies include the activation timing, identification accuracy requirements, or computing resource allocation methods. The system identifies and locates the person's hand area and segments the object being held. It then fuses synchronously collected multimodal data to identify the type of waste in the object. The identification of the type of waste in the object includes: extracting shape, material, texture, and structural features from the segmented area of ​​the object, combining the extracted hand posture features, voice and text data keywords, and the person's historical personality behavior, and using a pre-built machine learning model to identify the type of waste. The identified waste types are mapped to preset waste classification standards to generate recommended disposal types for the handheld object.

[0007] By adopting the above scheme, multi-source data around the garbage recycling bins are collected in real time. Different recognition strategies are triggered based on the entry of people into the zone and their action intentions, thus optimizing the use of computing resources. Feature extraction and multi-faceted feature fusion of handheld objects are performed to identify garbage types, improving the accuracy of garbage type identification. The identified garbage types are mapped to preset standards to generate recommended disposal types, improving the accuracy and efficiency of garbage classification and providing users with accurate disposal guidance.

[0008] Preferred options also include: A scene feature library is constructed to distinguish and store the behavioral characteristics of people in different scenes, and the confidence threshold for judging and predicting action intentions is adaptively adjusted by combining time, environment and location features. Based on real-time collection of multi-source data on human behavior, environment, time and location within a pre-divided area around the garbage recycling bin, the scene is determined by comparing it with a scene feature library, and confidence thresholds for judging and predicting action intent are matched.

[0009] By adopting the above scheme, and combining time, environment and location characteristics, the confidence threshold for judging and predicting action intentions is adjusted according to different scenarios to improve the accuracy of judging and predicting people's action intentions, thereby making the recommendation of waste disposal type more accurate.

[0010] Preferred options also include: During the triggering of different identification strategies, waste disposal identification algorithms are set to match different identification accuracy requirements, including: a first identification algorithm matching basic accuracy, a second identification algorithm matching medium accuracy, and a third identification algorithm matching high accuracy; the first identification algorithm uses a lightweight feature extraction network; the second identification algorithm uses a feature enhancement extraction algorithm and a scene environment weight adaptive adjustment algorithm; the third identification algorithm uses a multi-branch feature enhancement algorithm, multi-modal feature fusion, and scene environment weight adaptive adjustment algorithm.

[0011] By adopting the above scheme, matching waste disposal identification algorithms are set for different identification accuracy requirements, which can flexibly adapt to different scenarios and needs. While ensuring identification accuracy, computing resources are reasonably allocated to improve the efficiency and accuracy of waste type identification.

[0012] Preferred options also include: In the process of integrating synchronously collected multimodal data to identify the type of garbage in handheld objects, the motion state of the garbage in the captured handheld object is transformed into a quantifiable difficulty index, including: extracting the positional motion features, morphological change features, motion speed features, and adhesion state features of the garbage in the captured handheld object, and performing a weighted score based on the extracted features of the garbage in the captured handheld object to determine the quantitative result of the current difficulty level of garbage identification. Based on the current difficulty level quantification results of waste identification, the system determines the identification accuracy adjustment by identifying results that exceed the preset difficulty level threshold of the current segmented area. The adjusted identification accuracy level is then determined, and the corresponding algorithm enhancement module in the pre-built machine learning model is activated to optimize the model. The optimized machine learning model algorithm is then used to identify the type of waste in a handheld object. The pre-built machine learning model is based on a lightweight feature extraction network and multimodal feature fusion, and includes algorithm enhancement modules such as an attention module, a dynamic morphological correction module, a multi-branch feature enhancement module, a scene environment weight adaptive adjustment module, an instance segmentation algorithm, and an attribute analysis module. Different identification accuracy levels correspond to different preset algorithm enhancement modules to be activated.

[0013] By adopting the above scheme, the movement state of waste is transformed into a quantifiable difficulty indicator. Based on this, the recognition accuracy is adjusted and the machine learning model is optimized to adapt to different waste recognition difficulties and improve the accuracy of waste type recognition.

[0014] Preferred options also include: After identifying the waste type, a multi-factor collaborative quantitative score is calculated based on the preset risk level of the identified waste type, the consistency of waste classification across multiple time periods, the number of identified waste types, and the confidence level of the identified waste type, to obtain the quantitative score result. The quantitative score result triggers verification and matches the verification level accordingly, including a basic verification level that matches the quantitative score result within the first quantitative score range, an enhanced verification level that matches the quantitative score result within the second quantitative score range, and a deep verification level that matches the quantitative score result within the third quantitative score range. The identification of waste types is verified according to the matching verification level. Upon successful verification, the corresponding waste type is output; otherwise, the waste type identification is re-performed. The verification process for the basic verification level includes: cross-verifying the waste type identification results using redundant pre-built machine learning models. The verification process for the enhanced verification level includes: establishing multimodal data synchronously collected by multiple sets of sensors, fusing the multimodal data from each set, and using multiple pre-built machine learning models built into the edge computing center to output a set of waste type identification results, cross-verifying the accuracy of the waste type identification results. The verification process for the deep verification level includes: based on the enhanced verification results, designing an interactive interface for the waste bin to display the identified waste types, generating interactive confirmation options for waste type classification results, receiving user-input interactive confirmation results for waste type classification results, and cross-verifying the accuracy of the waste type identification results.

[0015] By adopting the above scheme, after identifying the waste type, the identification results are calculated using multi-factor collaborative quantitative scoring. Different levels of verification are triggered based on the scoring results, thereby effectively verifying the identification results and improving the accuracy of waste type classification. When the verification fails, the identification is re-performed, further ensuring the reliability of the final waste type classification output.

[0016] Preferred options also include: In the process of mapping the identified waste types to preset waste classification standards and generating recommended disposal types for the handheld object, multi-dimensional mapping rules are pre-set, and the waste classification standards are mapped according to the multi-dimensional mapping rules. The multi-dimensional mapping rules include: basic mapping rules that directly map according to the national standard waste classification standard library and the local standard waste classification standard library, and mapping corrections to the basic mapping results based on waste attributes and waste mixing types. Based on the waste type mapping results, a structured classification and disposal recommendation list is generated; the classification and disposal recommendation list contains each identified waste object and its recommended classification and disposal category, sorted according to the deterministic order of waste object identification; Based on real-time collection of multi-source data within a pre-divided area around the waste recycling bin, the current regional scenario and user age or personalized needs are determined, and a corresponding waste disposal recommendation strategy is output that is adapted to the current regional scenario and user age or personalized needs; the waste disposal recommendation strategy includes waste disposal recommendation content, auxiliary waste disposal content, and operation instructions for waste disposal.

[0017] By adopting the above scheme, the waste type is accurately mapped to the waste classification standard according to the multi-dimensional mapping rules, generating a structured and sorted classification recommendation list, and outputting an adapted disposal recommendation strategy according to the actual scenario and user needs, thereby improving the accuracy and personalization of waste classification recommendations.

[0018] Preferred options also include: When multiple handheld objects are identified and located, dynamic binding is performed between the person and the handheld object, and an association identifier is established; when the waste type of multiple handheld objects is identified, priority is determined according to the obtained quantitative score results, and the higher the quantitative score result, the higher the priority ranking; according to the determined priority ranking, the waste disposal recommendation strategy with the bound person information is displayed in an orderly manner.

[0019] By adopting the above scheme, when there are multiple handheld objects, users are bound to and sorted with the handheld objects, and the display priority is determined based on the quantitative scoring results, thereby improving the targeting and orderliness of user identification and recommendation, and thus optimizing the waste disposal recommendation process.

[0020] Secondly, this application provides a computer vision-based system for recommending types of household waste disposal, including: The data acquisition module is used to collect multi-source data in real time within a pre-defined area around the waste recycling bin using multimodal sensors; The trigger recognition module is used to detect and track people approaching the recycling bin based on the collected data, and to judge and predict the intention of the people's actions. When the tracking person approaches the recycling bin and enters a preset zone or the judgment and prediction of the person's intention to dispose of the waste is made based on the actions of the person, the recognition is triggered and different recognition strategies are triggered according to the different preset zones entered or the different confidence levels of the judgment and prediction of the disposal intention. The recognition strategy includes the activation time, recognition accuracy requirements or computing resource allocation method. The type recognition module is used to identify and locate the hand area of ​​the person and segment the handheld object; it integrates synchronously collected multimodal data to identify the type of waste in the handheld object; the identification of the waste type of the handheld object includes: extracting shape, material, texture and structural features from the segmented area of ​​the handheld object, combining the extracted hand posture features of the person, keywords of voice and text data and the person's historical personality behavior, and using a pre-built machine learning model to identify the waste type. The disposal recommendation module is used to map the identified waste types to preset waste classification standards and generate disposal type recommendation results for the handheld object.

[0021] By adopting the above solution, real-time data around the waste recycling bins is collected, and different recognition strategies are triggered based on the location and intention of the person to identify the type of waste being held and generate disposal recommendations, thereby improving the efficiency and accuracy of waste sorting. At the same time, the use of computing resources is optimized and the user experience is enhanced.

[0022] Thirdly, this application provides a computer-readable storage medium including a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to perform the method described above.

[0023] Fourthly, this application provides a computer device, the computer device including a memory, a processor and a program stored in the memory and executable thereon, the program being executed by the processor to implement the steps of the method described above.

[0024] In summary, this application has the following beneficial effects: 1. Real-time and comprehensive collection of multi-source data around the garbage recycling bins, real-time detection and tracking of personnel, triggering appropriate identification strategies based on personnel actions and their location to optimize resource utilization; combining multiple features and machine learning models to accurately identify garbage types, mapping the identified garbage types to preset standards, generating accurate and suitable disposal recommendations, and improving the efficiency and accuracy of garbage sorting. 2. Construct a scene feature library to adjust the confidence threshold for judging and predicting action intent; set different waste disposal recognition algorithms according to different recognition accuracy requirements; adjust the recognition accuracy and optimize the model based on the difficulty level of waste recognition determined by the movement state of the waste; perform multi-factor collaborative quantitative scoring and verification on the recognition results, generate a classification recommendation list and an appropriate disposal recommendation strategy according to multi-dimensional mapping rules, and dynamically bind and prioritize multiple handheld objects to display recommendation strategies; improve the accuracy and efficiency of waste classification and enhance the user experience. Attached Figure Description

[0025] Figure 1 This is a flowchart of the computer vision-based method for recommending household waste disposal types as described in a specific embodiment; Figure 2 This is a schematic diagram of the computer vision-based household waste disposal type recommendation system described in a specific embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] like Figure 1 As shown in the figure, this application discloses a method for recommending the type of household waste disposal based on computer vision, including steps such as data collection, trigger recognition, type recognition and disposal recommendation. The content of each step is described in detail below.

[0028] S1. Real-time collection of multi-source data within a pre-defined area around the waste recycling bin using multi-modal sensors.

[0029] Specifically, a multimodal sensing system incorporating multimodal sensors is pre-constructed to collect real-time data from a pre-defined area around the waste recycling bin, including image data, sound data, and environmental data. The construction process of the multimodal sensing system includes: First, deploy visual acquisition devices. Install visual acquisition devices in appropriate locations within the smart recycling bins to ensure clear capture of images of approaching people and their handheld items. Specifically, the visual acquisition devices include high-definition cameras and depth sensors. The high-definition camera acquires color images, providing visual features such as the color and texture of the waste; the depth sensor acquires 3D information of the scene, helping to separate foreground objects and determine the distance between residents and the bin. The camera can be mounted on the upper part of the bin, with a slightly downward tilted viewing angle, covering the approaching area in front of the bin. Second, deploy other sensor devices. Install other types of sensors in appropriate locations within the smart recycling bins, such as: sound sensors to collect sound data from a pre-defined area around the bin; and environmental data sensor arrays to collect environmental data from the pre-defined area around the bin.

[0030] Second, establish a sensing area model. Define the sensing area around the smart waste recycling bin to trigger the identification process and adjust the identification strategy. In this embodiment, the sensing area is pre-divided into a far-distance area (e.g., 3-5 meters in front of the waste bin), a medium-distance area (e.g., 1-3 meters in front of the waste bin), and a near-distance area (e.g., within 1 meter in front of the waste bin). The far-distance area detects residents approaching early and initiates identification preparation; the medium-distance area serves as the main waste identification area, where personnel entering this area begin to analyze the waste held by the resident; the near-distance area is the final confirmation area, where personnel arrive to complete the identification and provide the final recommendation.

[0031] Third, configure the vision processing parameters. Based on the actual application environment and recognition requirements, configure the relevant vision processing parameters. For example, set image acquisition parameters, including resolution, frame rate, and exposure settings, to ensure clear images are acquired under various lighting conditions; configure depth data processing parameters, including depth range, point cloud density, and filtering settings; set target detection parameters, including detection sensitivity, minimum target size, and tracking persistence; and define feature extraction parameters, including the number of feature points, descriptor type, and matching threshold. These parameters are initially set through the configuration interface and can be adjusted based on actual performance.

[0032] S2. Based on the collected data, detect and track people approaching the waste recycling bin, and perform judgment and prediction of people's action intentions to trigger waste type identification and determine the identification strategy.

[0033] First, personnel detection is performed. Specifically, the acquired color images and depth data are preprocessed, such as through noise filtering, illumination compensation, and depth filtering, to improve data quality. Person detection algorithms are then applied, using deep learning models such as convolutional neural networks or region proposal networks. These algorithms combine color images and depth information for multimodal fusion detection to identify personnel targets in the captured images. The detection results include the personnel's location, bounding boxes, and detection confidence scores.

[0034] Secondly, the detected personnel targets are continuously tracked, and their approach trajectories are recorded. A unique identifier is assigned to each detected person, and a target tracking model is established. Personnel targets are matched and updated in consecutive data frames to maintain identifier consistency. The tracking process employs a multi-feature fusion strategy, combining appearance features, depth information, and motion prediction, with particular attention to the personnel's direction of movement and speed, to continuously track personnel targets and record their approach trajectories.

[0035] Then, based on the approach trajectory and behavioral characteristics of personnel, the system analyzes and predicts their intention to approach the recycling bin. This involves tracking personnel approaching the bin into a pre-defined zone or judging and predicting their disposal intentions based on their actions, triggering waste type identification. Specifically, the judgment or prediction of personnel's intention to approach the recycling bin can utilize deep learning algorithms (such as lightweight LSTM networks). This analysis examines factors such as whether the personnel's movement direction is consistently towards the bin, whether their movement speed is suitable for disposal, whether their stopping position is in front of the bin, whether their posture changes are facing the bin, and whether they have preparatory actions for disposal. Features are extracted and comprehensively judged or predicted to determine the approach intention. Approaching personnel are categorized into those with disposal intention: high-intention disposal users (clearly intending to dispose of waste), potential disposal users (approaching but with unclear intention), and those without disposal intention: passersby (merely passing by without disposal intention). The system then judges or predicts whether disposal intention exists and its corresponding confidence level.

[0036] Finally, different identification strategies are triggered based on the different preset zones entered or the different confidence levels of the judged and predicted delivery intentions. These strategies include activation timing, required recognition accuracy, or computational resource allocation. Specifically, entering different preset zones includes: entering the mid-range area and the close-range area. Different confidence levels of the judged and predicted delivery intentions include: high confidence (…). 85% and medium confidence (60%-85%). Specifically, there are first and second identification strategies for the identification strategy; the first identification strategy, matching high confidence and medium / near distance areas, is set to start immediately, require high identification accuracy (greater than the first preset identification accuracy), and fully activate the edge NPU computing power; the second identification strategy, matching medium confidence and medium / near distance areas, is set to start late (e.g., monitor 20 frames), require medium identification accuracy (greater than the second preset identification accuracy), and enable 50% of the NPU computing power. In addition, for entering far-distance areas or low confidence ( The result of 60% does not trigger garbage type identification; it only continues monitoring but does not start identification analysis, thereby saving computing resources.

[0037] S3. Identify and locate the person's hand area, segment the object being held, fuse synchronously collected multimodal data, and identify the type of waste in the object being held.

[0038] Specifically, for individuals deemed to have the intention to dispose of waste, the focus is on locating their hand area to prepare for waste identification. Human keypoint detection technology is used beforehand to identify the skeletal structure and joint positions of the human body, accurately locating the wrist and palm areas. Then, the image and depth data of the hand area are finely processed, and an instance segmentation algorithm is applied to separate the hand from the held object, identifying the precise outline of the waste object. For complex scenarios, such as when multiple objects are held or objects are partially obscured by the hand, a temporal analysis method is used to supplement and infer the complete object outline through information from multiple consecutive frames. The segmentation results include: the mask, boundaries, and regional features of each held object, providing accurate target areas for subsequent object identification.

[0039] By integrating synchronously collected multimodal data, the system identifies the type of waste from handheld objects. Specifically, to further ensure the accuracy of waste type identification, a pre-built machine learning model is constructed to extract features from multiple dimensions, including visual, speech, and behavioral data, to identify the waste type and output the identification confidence score. Furthermore, when multiple handheld objects are input, the system outputs the waste type and its confidence score for each object. The pre-built machine learning model can employ a convolutional neural network, utilizing segmented image data of handheld objects from pre-divided areas around numerous recycling bins, speech data, and historical personal behavior data of individuals as input. It extracts shape, material, texture, and structural features from the segmented areas of the handheld objects, combining these with extracted hand posture features, speech text data keywords, and historical personal behavior features to output a waste type classification.

[0040] The extraction of hand gesture features includes: extracting the force of holding the trash (judged by the outline of arm muscles and the degree of finger bending), the speed of trash movement (judged by the change of trash position between time frames), and the preparatory actions for disposal (such as shaking the trash bag and splitting the trash); extracting keywords from the speech and text data includes: matching the text converted from speech with a trash classification keyword database (containing the standard names, common names, and material descriptions of more than 40 types of trash) to extract core keywords (such as batteries, plastic bottles, and leftovers), which are directly used as candidate results for recognition; and analyzing and extracting personal historical behavior features includes: combining locally stored user historical disposal data (anonymized) to extract personalized disposal habits (such as: a user often disposes of mixed trash such as takeout containers and leftovers, or a user prefers to use black trash bags to dispose of kitchen waste, establishing a personalized behavior-trash type association model).

[0041] Furthermore, based on the identification of waste types, the detailed attributes of the waste are further analyzed to provide a basis for accurate classification. This includes analyzing the material properties of the waste, such as plastic, paper, metal, glass, and organic matter; functional attributes, such as packaging materials, containers, food, and electronic products; and state attributes, such as clean, contaminated, damaged, and intact. These attribute analyses are achieved through feature extraction and pattern recognition, extracting the color distribution, texture features, reflective properties, and structural features of the waste and matching them with preset attribute models. The attribute analysis results form a multi-dimensional feature description of the waste, such as "plastic material, beverage container, intact state" or "paper material, food packaging, slightly contaminated," etc. Finally, the identification results are labeled, visually marking each identified waste object, including the object outline, type label, and confidence level. Different colors or label styles can also be used to distinguish individual handheld objects. The label information also includes the key attributes of the waste and the predicted classification category, such as recyclable plastic, kitchen waste, and hazardous waste.

[0042] S4. Map the identified waste type to the preset waste classification standard to generate a recommended disposal type result for the handheld object.

[0043] Specifically, it has a built-in waste sorting knowledge base, containing classification rules and special cases for common waste, based on national or local waste sorting standards. The mapping process includes: First, querying the direct classification rules for waste types, such as plastic bottles - recyclables, fruit peels - kitchen waste, and obtaining waste sorting standards based on the identified waste types; then, correcting the waste by analyzing its attributes, such as heavily soiled paper - other waste, rather than recyclables; in addition, for complex or ambiguous cases, rule reasoning and similar case matching can be used to find the most suitable classification category; finally, it outputs the final recommended disposal type for the handheld object; the final recommendation includes the main recommendation (recommended disposal port), auxiliary information (reason for disposal), and operation guidance (disposal order), using language descriptions, combined with visual labels and icons, and displayed to the user through the smart waste recycling bin.

[0044] In a specific embodiment, to further improve the accuracy of waste sorting and user experience, the method optimizes personnel detection in different scenarios to accurately trigger waste identification from the perspectives of triggering waste identification and adjusting waste identification strategies, and optimizes the identification algorithm under different waste identification strategies; the method also includes: To more accurately trigger waste identification, a scene feature library is constructed to distinguish differences in human behavior in different scenarios (e.g., residential areas, office buildings, and campuses). This includes different behaviors such as students carrying books on campus or people carrying documents in office buildings, and their hand gestures when holding books or documents. The library is then combined with time (e.g., 7-9 am or 6-8 pm), environmental (e.g., weather data), and location (latitude and longitude) features to adaptively adjust the confidence threshold for judging and predicting action intent. For example, in campus scenarios, 7-9 am or 6-8 pm are peak waste disposal times; under this feature combination, the confidence threshold for judging and predicting action intent is increased. In residential areas, in severe weather, this feature combination lowers the confidence threshold for judging and predicting action intent. Specifically, deep learning algorithms (e.g., lightweight LSTM networks) are used to extract features and comprehensively judge or predict near-intent. During this process, expert experience can be used for inspection or adaptive learning to adaptively adjust the confidence threshold for judging and predicting waste disposal intent under corresponding feature combinations, achieving more accurate waste disposal intent judgment.

[0045] Based on real-time collection of multi-source data on personnel behavior, environment, time and location within a pre-divided area around the waste recycling bin, the scene is determined by comparing it with a scene feature library (which stores corresponding features of personnel behavior, environment, time and location under different scenes), and a confidence threshold for judging and predicting the intention to dispose of waste is matched based on preset rules or by using deep learning algorithms.

[0046] In addition to triggering optimization for garbage identification, different identification algorithms can be set for different identification accuracy requirements. The identification strategy can be flexibly adjusted according to the actual situation, ensuring identification accuracy while optimizing the use of computing resources.

[0047] Specifically, during the triggering of different identification strategies, waste disposal identification algorithms are set to match different identification accuracy requirements, including: a first identification algorithm matching basic accuracy (e.g., accuracy threshold of the first threshold), a second identification algorithm matching medium accuracy (e.g., accuracy threshold of the second threshold), and a third identification algorithm matching high accuracy (e.g., accuracy threshold of the third threshold). The first identification algorithm uses a lightweight feature extraction network, such as a pruned YOLO algorithm. The second identification algorithm uses a feature enhancement extraction algorithm and a scene environment weight adaptive adjustment algorithm. Specifically, by improving the backbone network, it adopts a YOLOv12-Edge base, embeds dilated convolutions, and optimizes the attention mechanism, such as occlusion simulation enhancement, small target and complex position enhancement, and adaptively adjusts the weights of different types of features under different environmental data. The third identification algorithm uses a multi-branch feature enhancement algorithm, multi-modal feature fusion, and scene environment weight adaptive adjustment algorithm. Specifically, this includes: adding a high-resolution feature enhancement branch to directly extract high-resolution features from the backbone network, and supplementing the detailed features of small targets through upsampling and convolution fusion; when fusing multimodal features, different modalities can be fused separately (such as visual-speech dynamic fusion, pose-visual verification fusion) and then combined and weighted fusion.

[0048] In a specific embodiment, to further improve the accuracy and adaptability of handheld object waste type identification in different scenarios, the real-time captured handheld waste movement state is converted into a quantifiable difficulty index. The difficulty level of the current waste identification is accurately determined, and the identification accuracy is adjusted based on the difficulty level quantification result. The algorithm enhancement module is then activated to optimize the model, thereby adaptively adjusting according to the waste identification difficulty. The method also includes: In the process of integrating synchronously acquired multimodal data to identify the type of waste in handheld objects, the motion state of the waste in the captured handheld objects is transformed into quantifiable difficulty indicators. Specifically, this includes extracting the positional motion features, morphological change features, motion speed features, and adhesion status features of the waste in each captured handheld object. For positional motion features, inter-frame displacement and motion trajectory deviation can be calculated based on multi-view image frame matching and depth data containing historical image frames to determine whether the waste is shaking violently or deviating from the recognition area. For morphological change features, contour similarity (inter-frame), area change rate, and deformation degree (wrinkle / stretch level) can be obtained through contour extraction to determine whether the waste bag is deformed and the degree of deformation. For motion speed features, the average motion speed and instantaneous acceleration are calculated based on the displacement of N consecutive frames containing historical image frames to determine whether the disposal action is rapid. For adhesion status, the area change of the adhesion region is calculated based on the corresponding depth density difference and edge detection to determine whether the mixed waste separates during movement.

[0049] The difficulty level of garbage identification is determined by weighting and scoring various features of garbage extracted from each handheld object. For example, the weights are set as follows: morphological change 0.35, positional movement 0.3, movement speed 0.2, and adhesion state 0.15. Each feature range corresponds to a difficulty score, and a weighted comprehensive difficulty score (0-100 points) is calculated. The difficulty levels are divided into 3 levels: low, medium, and high.

[0050] Based on the current difficulty level quantification results of garbage identification, the system triggers an adjustment to the identification accuracy by determining if any result exceeds the preset difficulty level threshold of the currently (handheld object) segmented area. Specifically, the settings are as follows: If the current area is a mid-range area, and the difficulty level quantification results of garbage identification in the preceding frames are greater than medium difficulty for N consecutive frames, the identification accuracy is adjusted from basic accuracy to medium accuracy. If the subsequent N-1 consecutive frames return to low difficulty, the accuracy is reduced back to basic accuracy. If the current area is a mid-range area, and the difficulty level changes to increase for N consecutive frames including the preceding frames, the system triggers an adjustment supporting a full range of increases from basic accuracy to medium accuracy or from medium accuracy to high accuracy. A reduction in accuracy requires a decrease in difficulty level for N consecutive frames. If the current area is a near-range area, and the difficulty level of a single frame is medium difficulty or there is a change from low difficulty to medium difficulty for N consecutive frames including the preceding frames, the accuracy is adjusted upwards. A reduction in accuracy requires a stable low difficulty for N+1 consecutive frames.

[0051] The adjusted recognition accuracy level is determined, and the corresponding algorithm enhancement modules in the pre-built machine learning model are activated to complete model optimization. The optimized machine learning model algorithm is then used to identify the type of trash in a handheld object. Different adjusted recognition accuracy levels correspond to different preset algorithm enhancement modules to be activated. For example, after adjustment to base accuracy, the attention module and morphological dynamic correction module are dormant, and preset algorithm enhancement modules are not activated. After adjustment to medium accuracy, the multi-branch feature enhancement module and scene environment weight adaptive adjustment module are activated. Accordingly, the pre-built machine learning model is based on a lightweight feature extraction network and multimodal feature fusion, and algorithm enhancement modules are added, including an attention module, a morphological dynamic correction module, a multi-branch feature enhancement module, a scene environment weight adaptive adjustment module, an instance segmentation algorithm, and an attribute analysis module.

[0052] In one specific embodiment, multi-factor collaborative quantitative scoring and hierarchical verification are performed on the identified waste type classification to improve the accuracy and reliability of waste type identification and reduce misclassification. The method further includes: After identifying the waste types, a multi-factor collaborative quantitative scoring calculation is performed based on the preset risk level of the identified waste types, the consistency of waste classification across multiple time periods, the number of identified waste types, and the confidence level of the identified waste types, to obtain the quantitative scoring result. Specifically, the risk level of identified hazardous waste is preset as high risk, kitchen waste as medium risk, and other general waste as low risk, with corresponding quantitative scoring ranges set; the consistency of waste classification across multiple time periods includes: completely consistent, slightly inconsistent (number of inconsistent stages less than a preset threshold), and severely inconsistent (number of inconsistent stages not less than a preset threshold), with corresponding quantitative scoring ranges set; the number of identified waste types includes: single type, number that can be divided into multiple categories, and number that cannot be divided into multiple categories, with corresponding quantitative scoring ranges set; the confidence level of identified waste types includes: high, medium, and low, with corresponding quantitative scoring ranges set; and weights are assigned to each factor indicator, and a weighted calculation is performed to obtain the quantitative scoring result.

[0053] The verification is triggered and a verification level is matched according to the quantitative scoring result. This includes the basic verification level that matches the quantitative scoring result in the first quantitative scoring range (not less than 80 points), the enhanced verification level that matches the quantitative scoring result in the second quantitative scoring range (between 60 and 79 points), and the deep verification level that matches the quantitative scoring result in the third quantitative scoring range (less than 60 points). In other words, the higher the quantitative scoring range, the higher the reliability of the recognition result, and no complex verification is required.

[0054] Specifically, the identified garbage types are verified according to the matching verification level. If the verification is successful, the corresponding garbage type is output. If the verification fails, the garbage type is re-identified, such as after model optimization.

[0055] The verification process for the basic verification level includes: cross-validating the waste type identification results using redundant pre-built machine learning models, i.e., verifying the consistency between the original identification results and the redundant identification results. The verification process for the enhanced verification level includes: establishing multimodal data synchronously collected by multiple sets of sensors, fusing the multimodal data collected from each set, such as: handheld object image data and multimodal data corresponding to set A, handheld object image data and multimodal data corresponding to set B, etc. Multiple pre-built machine learning models built into the edge computing center are used to output a set of waste type identification results, and the accuracy of the waste type identification results is comprehensively verified (e.g., clustering verification or cross-validation). The verification process for the deep verification level includes: based on the enhanced verification results, integrating user feedback, designing an interactive interface for the waste bin to display the identified waste types, generating interactive confirmation options for waste type classification results, receiving user-input interactive confirmation results for waste type classification results, and comprehensively verifying (comparing the identified classification results with the user-confirmed classification results) the accuracy of the type classification identification results.

[0056] In a specific embodiment, to improve the accuracy and personalization of waste sorting recommendations and better meet the needs of different users and scenarios, a multi-dimensional mapping rule is used to accurately map waste types to waste sorting standards, generating a structured sorting recommendation list ordered by object salience and identifiability. An adapted disposal recommendation strategy is then output based on the actual scenario and user needs. The method further includes: First, in the process of mapping the identified waste types to preset waste classification standards and generating recommended disposal types for the handheld object, multi-dimensional mapping rules are pre-set. These rules are used to map the waste classification standards. The multi-dimensional mapping rules include: basic mapping rules that directly map to the national standard waste classification standard library (general standard library) and the local standard waste classification standard library (supplementary standard library), and mapping corrections based on waste attributes and waste mixing types. Therefore, in addition to corrections based on the acquired waste attributes, corrections can also be made based on waste mixing types. If multiple types of waste are identified in the handheld object, such as plastic bottles and leftover food, a separate mapping correction is performed.

[0057] Secondly, based on the waste type mapping results, a structured waste sorting recommendation list is generated. This list includes each identified waste object and its recommended sorting category, ordered according to the complexity and certainty of identification. Specifically, for single-type cases, the recommendation list is simpler, directly providing the waste type and its corresponding category. For multi-type cases, waste is grouped by category to form a category-oriented recommendation structure, such as: recyclables: plastic bottles, aluminum cans; kitchen waste: fruit peels, food scraps. The recommendation list also includes explanations for each sorting suggestion, such as: plastic bottles are recyclables, please put them in the blue recycling bin, to help users understand the rationale behind the recommendation. For waste with low identification confidence, multiple possible sorting options are provided, with a probability score for each option, for users to make a final judgment.

[0058] Based on real-time collection of multi-source data within a pre-defined area surrounding the recycling bin, the current regional scenario and user age or personalized needs are determined. Correspondingly, a waste disposal recommendation strategy is output, adapted to the current regional scenario and user age or personalized needs. This recommendation strategy includes recommended content, auxiliary content, and operational instructions for waste disposal. Specifically, if the current regional scenario is at a distance, only a simple welcome message is displayed; if the current regional scenario is at a medium distance, a preliminary recommended strategy is displayed; if the current regional scenario is at close range, a detailed recommended strategy is displayed, including main recommended content, auxiliary content, and operational instructions for waste disposal. During the generation of the recommended strategy, user age or personalized needs are further considered. For example, for users identified as younger by image recognition, a concise text and image format is used to display the recommended strategy; for older users, a video instruction format is provided, along with detailed step-by-step guidance. Furthermore, if the user enters their preferred display method, the recommended strategy is displayed accordingly.

[0059] Furthermore, when multiple handheld objects are identified and located, dynamic binding and association identification can be established between individuals and handheld objects to generate the final recommendation results in an orderly manner. For example, person A - handheld object 1 and handheld object 2, person B - handheld object 3, etc. When identifying the waste type of multiple handheld objects, priority is determined according to the obtained quantitative score results. The higher the quantitative score, the more accurate the current waste type identification, and the higher the priority ranking, the better the waste recommendation can be made. The waste recommendation strategies bound to the identified individuals are then displayed in an orderly manner according to the determined priority ranking, simultaneously displaying the individual's identification features, the associated handheld object image, and the waste recommendation strategy.

[0060] In a specific embodiment, in addition to considering waste type identification and generating disposal recommendations before waste disposal, after waste disposal, images of the waste can be re-captured based on the visual device inside the waste recycling bin, and multimodal data such as voice data and environmental data can be integrated to re-identify the waste type and check whether it is consistent with the waste classification standard corresponding to the waste recycling bin. If there is a discrepancy, a correction instruction is generated and prompted in the form of a voice broadcast; the method also includes: After generating the recommended disposal type of the handheld object, the system collects data on the disposed waste in real time and uses a pre-built machine learning model to re-identify the waste type. If the waste type does not belong to the preset waste classification standard corresponding to the current waste recycling bin, such as identifying hazardous waste as kitchen waste instead of kitchen waste, and if the system detects that the person associated with the waste disposal is still in the pre-defined area around the current waste recycling bin, it generates a correction instruction to remind the person who has not left to correct the disposal.

[0061] like Figure 2 As shown in the figure, this application discloses a computer vision-based system for recommending types of household waste disposal, specifically including: The data acquisition module 100 is used to collect multi-source data in real time within a pre-defined area around the waste recycling bin using a multimodal sensor; The trigger identification module 200 is used to detect and track people approaching the recycling bin based on the collected data, and to judge and predict the intention of the people's actions; when tracking people approaching the recycling bin to enter a preset zone or judging and predicting the intention of the people to dispose of the waste based on their actions, the module triggers identification and triggers different identification strategies based on the different preset zones entered or the different confidence levels of the judged and predicted intention to dispose of the waste; the identification strategy includes the activation time, identification accuracy requirements or computing resource allocation method; The type recognition module 300 is used to identify and locate the hand area of ​​the person and segment the handheld object; it integrates synchronously collected multimodal data to identify the type of waste in the handheld object; the identification of the waste type of the handheld object includes: extracting shape, material, texture and structural features from the segmented area of ​​the handheld object, combining the extracted hand posture features of the person, keywords of voice and text data and the person's historical personality behavior, and using a pre-built machine learning model to identify the waste type. The disposal recommendation module 400 is used to map the identified waste type to a preset waste classification standard and generate a disposal type recommendation result for the handheld object.

[0062] In one specific embodiment, the system further includes: The trigger recognition optimization module 500 is used to build a scene feature library to distinguish and store the behavioral characteristics of people in different scenes, and adaptively adjust the confidence threshold for judging and predicting action intentions by combining time, environment and location features. Based on the real-time collection of multi-source data on human behavior, environment, time and location in the pre-divided area around the garbage recycling bin, the scene feature library is compared to determine the scene, and the confidence threshold for judging and predicting action intentions is matched.

[0063] The trigger recognition optimization module 500 is also used to set up matching garbage disposal recognition algorithms for different recognition accuracy requirements during the triggering of different recognition strategies, including: a first recognition algorithm matching basic accuracy, a second recognition algorithm matching medium accuracy, and a third recognition algorithm matching high accuracy; the first recognition algorithm adopts a lightweight feature extraction network; the second recognition algorithm adopts a feature enhancement extraction algorithm and a scene environment weight adaptive adjustment algorithm; the third recognition algorithm adopts a multi-branch feature enhancement algorithm, multi-modal feature fusion, and scene environment weight adaptive adjustment algorithm.

[0064] In one specific embodiment, the system further includes: The type recognition optimization module 600 is used to identify the type of waste in a handheld object during the process of fusing synchronously collected multimodal data. It captures the motion state of the waste in the handheld object and converts it into quantifiable difficulty indicators. This includes extracting the positional motion features, morphological change features, motion speed features, and adhesion features of the waste in the handheld object. Based on these extracted features, a weighted score is applied to determine the current difficulty level of waste identification. According to the current difficulty level, if a result exceeds the preset difficulty level threshold for the current segmented area, an adjustment to the recognition accuracy is triggered. The adjusted recognition accuracy level is determined, and the corresponding algorithm enhancement module in the pre-built machine learning model is activated to optimize the model. The optimized machine learning model algorithm is then used to identify the type of waste in the handheld object.

[0065] In one specific embodiment, the system further includes: The type identification and verification module 700 is used to perform multi-factor collaborative quantitative scoring calculation after identifying and acquiring the waste type, based on the preset risk level of the identified waste type, the consistency of waste classification identified in multiple time periods, the number of identified waste types, and the confidence level of the identified waste type, to obtain a quantitative score result; trigger verification and match the verification level according to the quantitative score result; verify the identified waste type according to the matched verification level, output the corresponding waste type after successful verification, and re-identify the waste type after failure.

[0066] In one specific embodiment, the system further includes: The waste disposal recommendation optimization module 800 is used to pre-set multi-dimensional mapping rules during the process of mapping the identified waste types to preset waste classification standards and generating waste disposal type recommendation results for the handheld object. These multi-dimensional mapping rules include: basic mapping rules that directly map to national and local waste classification standard libraries, and mapping corrections to the basic mapping results based on waste attributes and waste mixing types; generating a structured waste disposal recommendation list based on the waste type mapping results; and determining the current area scenario and user age or personalized needs based on real-time collected multi-source data within a pre-divided area around the waste recycling bin, and outputting a corresponding waste disposal recommendation strategy adapted to the current area scenario and user age or personalized needs.

[0067] The waste disposal recommendation optimization module 800 is also used to dynamically bind personnel to handheld objects and establish associated identifiers when multiple handheld objects are identified and located; when the waste type of multiple handheld objects is identified, priority is determined according to the obtained quantitative score results, and the higher the quantitative score result, the higher the priority ranking; and the waste disposal recommendation strategy bound with personnel information is displayed in an orderly manner according to the determined priority ranking.

[0068] This application also discloses a computer-readable storage medium.

[0069] Specifically, the computer-readable storage medium stores a computer program that can be loaded by a processor and executed, such as the computer vision-based method for recommending household waste disposal types described above. The computer-readable storage medium includes, for example, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] This application also discloses a computer device.

[0071] Specifically, the computer device includes a memory and a processor, and the memory stores a computer program that can be loaded by the processor and executed to recommend the type of household waste disposal based on computer vision.

[0072] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A computer vision-based method for recommending types of household waste disposal, characterized in that, include: Multi-source data is collected in real time within a pre-defined area around the waste recycling bin using multi-modal sensors; Based on the collected data, detect and track people approaching the recycling bins, and judge and predict the intentions of people's actions; The system tracks individuals approaching waste recycling bins and entering preset zones, or determines and predicts their disposal intentions based on their movements. It then triggers identification strategies based on the confidence levels of individuals entering different preset zones or their predicted disposal intentions. These identification strategies include activation timing, identification accuracy requirements, and computational resource allocation methods. Identify and locate the person's hand area, and segment the object being held; By integrating synchronously collected multimodal data, the type of waste in handheld objects can be identified; The method of identifying the type of waste from handheld objects includes: extracting shape, material, texture and structural features from the segmented region of the handheld object, combining the extracted hand posture features, voice and text data keywords and the person's historical personality behavior, and using a pre-built machine learning model to identify the type of waste. The identified waste types are mapped to preset waste classification standards to generate recommended disposal types for the handheld object.

2. The method for recommending household waste disposal types based on computer vision according to claim 1, characterized in that, Also includes: A scene feature library is constructed to distinguish and store the behavioral characteristics of people in different scenes, and the confidence threshold for judging and predicting action intentions is adaptively adjusted by combining time, environment and location features. Based on real-time collection of multi-source data on human behavior, environment, time and location within a pre-divided area around the garbage recycling bin, the scene is determined by comparing it with a scene feature library, and confidence thresholds for judging and predicting action intent are matched.

3. The method for recommending household waste disposal types based on computer vision according to claim 1, characterized in that, Also includes: During the triggering of different identification strategies, waste disposal identification algorithms are set to match different identification accuracy requirements, including: a first identification algorithm matching basic accuracy, a second identification algorithm matching medium accuracy, and a third identification algorithm matching high accuracy; the first identification algorithm adopts a lightweight feature extraction network; the second identification algorithm adopts a feature enhancement extraction algorithm and a scene environment weight adaptive adjustment algorithm; the third identification algorithm adopts a multi-branch feature enhancement algorithm, multi-modal feature fusion, and scene environment weight adaptive adjustment algorithm.

4. The method for recommending household waste disposal types based on computer vision according to claim 1, characterized in that, Also includes: In the process of integrating synchronously collected multimodal data to identify the type of garbage in handheld objects, the motion state of the garbage in the captured handheld object is transformed into a quantifiable difficulty index, including: extracting the positional motion features, morphological change features, motion speed features, and adhesion state features of the garbage in the captured handheld object, and performing a weighted score based on the extracted features of the garbage in the captured handheld object to determine the quantitative result of the current difficulty level of garbage identification. Based on the current difficulty level quantification results of waste identification, the system determines the identification accuracy adjustment by identifying results that exceed the preset difficulty level threshold of the current segmented area. The adjusted identification accuracy level is then determined, and the corresponding algorithm enhancement module in the pre-built machine learning model is activated to optimize the model. The optimized machine learning model algorithm is then used to identify the type of waste in a handheld object. The pre-built machine learning model is based on a lightweight feature extraction network and multimodal feature fusion, and includes algorithm enhancement modules such as an attention module, a dynamic morphological correction module, a multi-branch feature enhancement module, a scene environment weight adaptive adjustment module, an instance segmentation algorithm, and an attribute analysis module. Different identification accuracy levels correspond to different preset algorithm enhancement modules to be activated.

5. The computer vision-based method for recommending household waste disposal types according to claim 4, characterized in that, Also includes: After identifying the waste type, a multi-factor collaborative quantitative scoring calculation is performed based on the preset risk level of the identified waste type, the consistency of waste classification identified in multiple time periods, the number of identified waste types, and the confidence level of the identified waste type to obtain the quantitative scoring result. Based on the quantitative scoring results, corresponding verification is triggered and the verification level is matched, including the basic verification level that matches the quantitative scoring results in the first quantitative scoring range, the enhanced verification level that matches the quantitative scoring results in the second quantitative scoring range, and the deep verification level that matches the quantitative scoring results in the third quantitative scoring range. The identified garbage type is verified according to the matching verification level. If the verification is successful, the corresponding garbage type is output. If the verification fails, the garbage type is re-identified. The basic verification level verification process includes: cross-verifying the waste type identification results using redundant pre-built machine learning models; the enhanced verification level verification process includes: establishing multimodal data collected synchronously by multiple sets of sensors, fusing the multimodal data collected synchronously by each set, and using multiple pre-built machine learning models built into the edge computing center to output a set of waste type identification results, cross-verifying the accuracy of the waste type identification results; the deep verification level verification process includes: based on the enhanced verification results, designing a waste bin interactive interface to display the identified waste types, generating interactive confirmation options for waste type classification results, receiving user-input interactive confirmation results for waste type classification results, and cross-verifying the accuracy of the waste type identification results.

6. The method for recommending household waste disposal types based on computer vision according to claim 1, characterized in that, Also includes: In the process of mapping the identified waste types to preset waste classification standards and generating recommended disposal types for the handheld object, multi-dimensional mapping rules are pre-set, and the waste classification standards are mapped according to the multi-dimensional mapping rules. The multi-dimensional mapping rules include: basic mapping rules that directly map according to the national standard waste classification standard library and the local standard waste classification standard library, and mapping corrections to the basic mapping results based on waste attributes and waste mixing types. Based on the waste type mapping results, a structured classification and disposal recommendation list is generated; the classification and disposal recommendation list contains each identified waste object and its recommended classification and disposal category, sorted according to the deterministic order of waste object identification; Based on real-time collection of multi-source data within a pre-divided area around the waste recycling bin, the current regional scenario and user age or personalized needs are determined, and a corresponding waste disposal recommendation strategy is output that is adapted to the current regional scenario and user age or personalized needs; the waste disposal recommendation strategy includes waste disposal recommendation content, auxiliary waste disposal content, and operation instructions for waste disposal.

7. The computer vision-based method for recommending household waste disposal types according to claim 5, characterized in that, Also includes: When multiple handheld objects are identified and located, dynamic binding between the person and the handheld object is performed and an association identifier is established; The system identifies the types of waste from multiple handheld objects and prioritizes them according to the obtained quantitative scores. The higher the quantitative score, the higher the priority. Based on the determined priority, a waste disposal recommendation strategy that is linked to personnel information is displayed in an orderly manner.

8. A computer vision-based system for recommending types of household waste disposal, characterized in that, include: The data acquisition module is used to collect multi-source data in real time within a pre-defined area around the waste recycling bin using multimodal sensors; The trigger recognition module is used to detect and track people approaching the recycling bin based on the collected data, and to judge and predict the intentions of the people's actions. The system tracks individuals approaching waste recycling bins and entering preset zones, or determines and predicts their disposal intentions based on their movements. It then triggers identification strategies based on the confidence levels of individuals entering different preset zones or their predicted disposal intentions. These identification strategies include activation timing, identification accuracy requirements, and computational resource allocation methods. The type recognition module is used to identify and locate the hand area of ​​the person and segment the handheld object; By integrating synchronously collected multimodal data, the type of waste in handheld objects can be identified; The method of identifying the type of waste from handheld objects includes: extracting shape, material, texture and structural features from the segmented region of the handheld object, combining the extracted hand posture features, voice and text data keywords and the person's historical personality behavior, and using a pre-built machine learning model to identify the type of waste. The disposal recommendation module is used to map the identified waste types to preset waste classification standards and generate disposal type recommendation results for the handheld object.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device on which the computer-readable storage medium is located to perform the method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, The computer device includes a memory, a processor, and a program stored in and executable on the memory, the program being executed by the processor to implement the steps of the method as described in any one of claims 1 to 7.

Images