High-precision lightweight classification detection method and device based on multiple types of garbage

Abstract

The invention provides a high-precision lightweight classification detection method and device based on multiple types of garbage. The device comprises a garbage target detection model, an automatic detection structure and a garbage sorting structure. The junk target detection model is an improved lightweight computing platform based on YOLOv11, and the junk target detection model is characterized in that GhostConv is integrated in a feature extraction link of YOLOv11, and BiFPN is integrated in a feature fusion link; the automatic detection structure is used for transporting garbage and collecting garbage images, and assists the lightweight computing platform in classifying and identifying the garbage; and the garbage sorting structure receives the decision instruction of the garbage target detection model and executes a sorting action. While high detection precision is maintained, calculation and storage requirements are greatly reduced, new garbage categories can be adapted by replacing training data, a hardware production line does not need to be changed, multi-scene detection of streets, dark environments, small targets, targets difficult to recognize and the like can be adapted, and garbage classification accuracy is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of garbage classification, and specifically provides a high-precision lightweight classification detection method and device based on multi-class garbage. BACKGROUND

[0002] With the acceleration of urbanization, garbage classification has become a key challenge in the field of environmental protection. The core goal of garbage classification is to separate different components in mixed waste for subsequent processing such as recycling, incineration, landfill, and fertilization.

[0003] The original garbage classification completely relies on manual sorting, which is inefficient and inaccurate, and the harsh working environment endangers the health of workers. Moreover, with the disappearance of the demographic dividend and the rising labor cost, manual sorting is costly. The mainstream industrialized garbage classification technology that emerged to overcome the shortcomings of manual sorting mainly uses the macro-physical properties of garbage such as size, weight, magnetism, density, and optical properties for sorting. The mainstream industrialized technology is a one-size-fits-all approach that cannot make fine distinctions among different items in the same material category; it requires high pretreatment, and the equipment usually needs the garbage to be pretreated by crushing and loosening; it has poor flexibility, and once the production line is built, it is difficult to adapt to changes in garbage composition or new sorting categories.

[0004] Currently, the mainstream garbage sorting method mainly relies on manual operation or traditional automatic detection systems, but these methods have obvious limitations in resource consumption, deployment flexibility, and precision. SUMMARY

[0005] To solve the above problems, in a first aspect, the present application provides a high-precision lightweight classification detection device based on multi-class garbage, which includes a garbage target detection model, an automatic detection structure, and a garbage sorting structure. The garbage target detection model is a lightweight computing platform based on the improvement of YOLOv11. The garbage target detection model integrates GhostConv in the feature extraction link of YOLOv11 and integrates BiFPN in the feature fusion link. The automatic detection structure is used to transport garbage and collect garbage images, assisting the lightweight computing platform in classifying and identifying garbage. The garbage sorting structure receives the decision instruction of the garbage target detection model and performs a sorting action.

[0006] In one technical solution of the above-mentioned high-precision lightweight classification detection device based on multi-class garbage, the GhostConv filters the feature maps of multi-class garbage by setting a feature response threshold, retains high-response features corresponding to the texture and shape profile of the garbage material, and eliminates low-response features corresponding to stains and shadows.

[0007] In one technical solution of the above-mentioned high-precision lightweight classification and detection device based on multiple types of waste, the dynamic weight allocation strategy of BiFPN is as follows: through trainable weight parameters, the feature layer corresponding to small target waste is assigned a higher weight than the conventional target feature layer, and the weight parameters are obtained by training with samples of multiple types of waste.

[0008] In one technical solution of the above-mentioned high-precision lightweight classification and detection device based on multiple types of waste, the training dataset of the waste target detection model is cleaned and adapted, and contains at least 20 types of common household waste. The data cleaning includes correcting the overlapping labeled samples caused by waste stacking.

[0009] In one technical solution of the above-mentioned high-precision lightweight classification and detection device based on multiple types of waste, the data cleaning also includes supplementing waste scene samples with stains, deformation and damage to obtain a high-quality training dataset.

[0010] In one technical solution of the above-mentioned high-precision lightweight classification and detection device based on multiple types of waste, the automated detection structure includes a conveyor belt and a camera. The transport speed of the conveyor belt is matched with the end-to-end delay of the waste target detection model, and the installation height and resolution of the camera are adapted to image acquisition in the waste stacking scenario.

[0011] In one technical solution of the aforementioned high-precision lightweight classification and detection device for multiple types of waste, the image acquisition method for adapting the camera's installation height and resolution to the waste stacking scenario specifically includes: the camera adopts an oblique overhead shooting installation method, is configured with an adjustable focal length to adapt to the height range of the waste stacking, and establishes a connection with the computing platform through a high-speed data interface; the high-speed data interface is configured with a dynamic bandwidth allocation mechanism and a feature packet transmission protocol, and encapsulates and marks the location information of the acquired image data according to the upper layer exposed features and the lower layer edge features; the high-speed data interface maintains real-time interaction with the computing platform, and when the computing platform detects that the lower layer edge feature data is incomplete, it triggers the camera to re-shoot the target area. After receiving the complete data, the computing platform synchronously performs layered feature pre-labeling and edge enhancement preprocessing operations.

[0012] In one technical solution of the above-mentioned high-precision lightweight classification and detection device based on multiple types of waste, the waste sorting structure includes an execution mechanism and a sensor. The execution mechanism controls the movement of the robotic arm and the end effector through a controller. The sensor is used to assist in the positioning of waste and works in conjunction with the automated detection structure and the waste target detection model to form a detection-sorting closed loop.

[0013] In one technical solution of the aforementioned high-precision lightweight classification and detection device based on multiple types of waste, the waste target detection model is equipped with an image preprocessing module and a feature enhancement processing logic. The image preprocessing module uses an illumination compensation algorithm to adjust the brightness and contrast of waste images collected in dim street environments. The feature enhancement processing logic first enlarges the features of small and difficult-to-identify waste targets, then strengthens the target edges by adjusting the contrast of feature channels, and finally performs feature highlighting processing based on the grayscale differences of texture features to adapt to the detection requirements of relevant targets.

[0014] Secondly, this invention provides a high-precision, lightweight classification and detection method for multiple types of waste, comprising the following steps: S1, waste is transported via a conveyor belt of an automated detection structure, and a camera simultaneously acquires waste images and transmits them to a waste target detection model; S2, the waste target detection model first adjusts the image brightness and contrast through an image preprocessing module, and then performs feature extraction and classification recognition on the waste images through the synergistic effect of BiFPN trainable weight allocation and GhostConv feature filtering, outputting classification decision results; S3, the waste sorting structure receives the classification decision results, uses sensors for assisted positioning, and the execution mechanism drives the robotic arm and end effector to complete waste sorting.

[0015] The beneficial effects of the present invention are as follows: 1. The garbage target detection model of the present invention significantly reduces the computational and physical storage requirements of the model while maintaining high detection accuracy by making lightweight improvements to YOLOv11 based on BiFPN and GhostConv.

[0016] 2. By changing the training data, it can learn and sort new types of waste without changing the hardware production line. It can achieve waste detection in multiple environments, including various typical scenarios such as street environments, dimly lit conditions, small targets, and difficult-to-identify targets.

[0017] 3. By combining mechanical structures, manual intervention can be reduced and the level of automation in waste sorting can be improved. Attached Figure Description

[0018] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This is a network structure diagram of a garbage target detection model according to an embodiment of the present invention. Detailed Implementation

[0019] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0020] Example 1 like Figure 1 As shown, this invention provides a high-precision, lightweight classification and detection device for multiple types of waste, including: a waste target detection model, an automated detection structure, and a waste sorting structure. The waste target detection model is a lightweight computing platform based on an improved version of YOLOv11. This model integrates GhostConv in the feature extraction stage of YOLOv11 and BiFPN in the feature fusion stage, significantly reducing the computational complexity and number of parameters while maintaining high detection accuracy. This allows the model to better adapt to the practical application needs in resource-constrained environments. Specifically, the Ghost convolution module effectively reduces computation by generating redundant feature maps; the BiFPN structure enhances multi-scale feature fusion capabilities, particularly helping to improve the detection accuracy of small-target waste. Experimental verification shows that the optimized model significantly reduces the number of parameters and computation while maintaining high performance, making it more suitable for deployment on edge devices or in resource-constrained environments, greatly improving the model's operating efficiency and deployment convenience. The automated detection structure is used to transport waste and collect waste images, assisting the lightweight computing platform in classifying and identifying waste; the waste sorting structure receives decision instructions from the waste target detection model and executes sorting actions.

[0021] In this embodiment, a publicly available dataset was collected from the Roboflow website. This dataset contains common garbage data from everyday life scenarios. After acquiring this data, to ensure data quality, images with incorrect annotations and mixed instance segmentation annotation types were manually corrected. After a series of rigorous data cleaning and filtering processes, a high-quality dataset was successfully constructed. This dataset contains 52,861 images covering 20 garbage categories, providing a solid and reliable data foundation for subsequent model training. In the automated detection structure, the camera is installed 1.2 meters above the conveyor belt, covering the entire working area.

[0022] In one embodiment, GhostConv filters feature maps of multiple types of waste by setting a feature response threshold, retaining high-response features corresponding to waste material texture and shape contour, and removing low-response features corresponding to stains and shadows.

[0023] In this embodiment, GhostConv first performs feature response analysis on the input waste image feature map (including various information such as material texture, shape contour, stains, and shadows), calculating a response value for each feature point. The material texture of the waste (e.g., the smoothness of plastic, the wrinkles of paper, the reflectivity of metal) and the shape contour (e.g., the cylindrical shape of a bottle, the square shape of a box) are core recognition features, corresponding to higher response values; while noise information unrelated to waste category judgment, such as stains and shadows, corresponds to lower response values. GhostConv can adaptively adjust a preset feature response threshold through experimental calibration or model training. This threshold serves as a selection criterion: pixel regions with feature response values ​​higher than the threshold (i.e., high-response features corresponding to waste material texture and shape contour) are retained, while pixel regions with feature response values ​​lower than the threshold are set to invalid values, i.e., low-response features corresponding to stains and shadows are removed. Through this selection process, GhostConv can remove irrelevant noise and strengthen core recognition features while maintaining lightweight computation, allowing subsequent model layers to focus on the key features of waste for classification and recognition, thus improving recognition accuracy and reducing computational overhead.

[0024] In one embodiment, the dynamic weight allocation strategy of BiFPN is as follows: by using trainable weight parameters, the feature layer corresponding to small target garbage is assigned a higher weight than the feature layer of regular target garbage, and the weight parameters are obtained by training multiple types of garbage samples.

[0025] In this embodiment, during feature fusion, higher weights are assigned to the feature layers corresponding to small-sized waste targets, allowing the model to focus more on these weak signal features during computation. Simultaneously, adaptive weights are assigned to the feature layers of regular-sized targets to balance the recognition needs of targets at different scales. These weight parameters are not fixed values ​​but are obtained through training on a large-scale dataset of multi-class waste samples (covering small targets, regular targets, and various waste types). During training, the model adaptively adjusts the weight parameters of each feature layer based on the recognition errors of targets at different scales, ultimately forming an optimized allocation scheme where "the feature layer weight for small targets is higher than that for regular targets." This ensures that the features of small-sized waste targets are not weakened after fusion, significantly improving their recognition accuracy, while also considering the recognition performance of regular targets.

[0026] In one embodiment, the training dataset of the garbage target detection model is cleaned and adapted to include at least 20 types of common household waste, and the data cleaning includes correcting overlapping labeled samples caused by garbage stacking.

[0027] This embodiment covers at least 20 types of common household waste (such as plastics, paper, metals, kitchen waste, glass, etc.), encompassing the mainstream types of daily waste. This allows the model to learn rich category features and improve its generalization ability. Because waste is easily squeezed and covered during actual dumping, the annotation boxes or regions of different waste types overlap during data annotation. If left unaddressed, this can lead to the model learning incorrect feature associations (such as confusing the boundaries between different waste categories). During data cleaning, samples with overlapping annotations are identified through manual verification or automated algorithms (such as cross-validation and boundary adjustment). These overlapping annotations are then split and corrected: either by adjusting the position and size of the annotation boxes to clearly distinguish the boundaries of different waste types, or by removing or re-annotating samples that are completely overlapping and cannot be split. This ensures that the annotation information for each waste target is accurate and independent, ultimately resulting in a comprehensive, accurately labeled training dataset free of invalid interference. This allows the model to effectively learn the independent features of various types of waste, improving detection accuracy.

[0028] In one embodiment, the data cleaning further includes supplementing garbage scene samples with stains, deformation, and damage to obtain a high-quality training dataset.

[0029] In this embodiment, in actual waste disposal scenarios, waste often exhibits stains (such as plastic boxes contaminated with kitchen waste or paper stained with oil) and deformation / damage (such as crushed plastic bottles or torn packaging bags). These samples differ significantly from intact, clean waste. If such samples are missing from the training data, the model will experience recognition errors due to the lack of learned features. Therefore, the data cleaning stage specifically supplements these special scenario samples: a large number of stained and deformed waste images are acquired through on-site collection or manual simulation. After being labeled with categories and boundaries according to a unified standard, these images are included in the training dataset. The supplemented dataset covers both regular, intact waste samples and waste samples in various special states, forming high-quality training data.

[0030] In one embodiment, the automated detection structure includes a conveyor belt and a camera, wherein the conveyor belt's transport speed is matched to the end-to-end delay of the waste target detection model, and the camera's installation height and resolution are adapted to image acquisition in the waste stacking scenario.

[0031] In this embodiment, the conveyor belt speed (e.g., 0.05 m / s) is precisely matched with the 40 ms end-to-end delay of the waste target detection model. This ensures that within the time it takes for the model to complete one inference cycle, the waste only moves slightly (e.g., only 2 mm) and remains within the camera's field of view, avoiding missed detections or image blurring due to excessively rapid movement. The camera's installation height is designed for scene adaptation, and with a 1920×1080 color RGB resolution and a 512×424 depth map resolution, it can fully cover the width of the conveyor belt and the height of the waste stack. Furthermore, the high pixel density clearly captures the outline, texture, and three-dimensional spatial information of the stacked waste. Even in scenarios with overlapping waste and varying heights, it provides high-quality visual data for the model, ensuring detection accuracy.

[0032] In one embodiment, the image acquisition method of adapting the camera's installation height and resolution to the garbage stacking scenario specifically includes: the camera is installed at an angle and shooting downwards, configured with an adjustable focal length to adapt to the height range of the garbage stacking, and establishes a connection with the computing platform through a high-speed data interface; the high-speed data interface is configured with a dynamic bandwidth allocation mechanism and a feature packet transmission protocol, and encapsulates the acquired image data according to the upper layer exposed features and the lower layer edge features, and marks the location information respectively; the high-speed data interface maintains real-time interaction with the computing platform, and when the computing platform detects that the lower layer edge feature data is incomplete, it triggers the camera to re-shoot the target area, and after the computing platform receives the complete data, it synchronously performs layered feature pre-labeling and edge enhancement preprocessing operations.

[0033] In this embodiment, the camera can be installed at an angle, shooting downwards. With an adjustable focal length, it can cover the entire width of the conveyor belt from a top-down perspective, and the focal length can be flexibly adapted to different waste stacking heights (e.g., adjusting the focal length closer when the stack is high and adjusting it further when the stack is low). This avoids some waste from exceeding the field of view or becoming blurry due to height differences. Simultaneously, the angled angle reduces shading from stacked waste, improving image clarity. The camera connects to the computing platform via a high-speed data interface. The interface is configured with a dynamic bandwidth allocation mechanism (e.g., adjusting transmission bandwidth in real time based on data volume to avoid congestion) and a feature packet transmission protocol. The acquired image data is encapsulated according to upper-layer exposed features (e.g., visible contours and textures on the surface of stacked waste) and lower-layer edge features (e.g., the edge contours of the bottom layer or overlapping areas of the stack), and their respective position information in the image is marked. This makes data transmission more targeted and improves processing efficiency. The high-speed data interface maintains real-time interaction with the computing platform. When the computing platform receives data, it verifies the integrity of the lower-layer edge features. If edge data is missing at the bottom layer or overlapping area of ​​the stacked garbage (which is easily caused by occlusion), the camera will be immediately triggered to re-capture the target area to ensure that no key features are missed. After the re-capture is completed, the computing platform receives the complete data and simultaneously performs hierarchical feature pre-labeling (marking the positions of upper and lower layer features in advance to facilitate rapid model localization) and edge enhancement preprocessing (enhancing edge details and reducing the impact of stacked occlusion on recognition), providing high-quality, multi-dimensional visual data support for the subsequent garbage detection model.

[0034] In one embodiment, the waste sorting structure includes an execution mechanism and sensors. The execution mechanism controls the movement of the robotic arm and end effector through a controller. The sensors are used to assist in the positioning of waste and work in conjunction with the automated detection structure and the waste target detection model to form a detection-sorting closed loop.

[0035] In one embodiment, the garbage target detection model includes an image preprocessing module and a feature enhancement processing logic. The image preprocessing module uses an illumination compensation algorithm to adjust the brightness and contrast of garbage images collected in dimly lit street environments. The feature enhancement processing logic first enlarges the features of small garbage targets and difficult-to-identify garbage targets, then strengthens the target edges by adjusting the contrast of feature channels, and finally performs feature highlighting processing based on the grayscale differences of texture features to adapt to the detection requirements of relevant targets.

[0036] In this embodiment, the illumination compensation algorithm adaptively adjusts the brightness and contrast of the image to improve the visibility of details in dark areas (such as making the outlines and textures of garbage in dim environments clear), and avoids the model being unable to capture effective features due to insufficient illumination.

[0037] The features are enhanced through a three-step process: First, feature scaling is performed: small target waste has a small proportion and weak signal in the image. Scaling techniques (such as interpolation and feature upsampling) are used to expand its feature dimension and prevent it from being masked by other target features. Second, feature channel contrast is adjusted: to address the problem of blurred target edges, the contrast between feature channels is enhanced to strengthen the edge contour of the waste (such as making the edge of deformed plastic bottles clearer) and clarify the target boundary. Third, texture features are highlighted based on gray-level differences: the gray-level distribution differences of different waste materials (such as the different gray-level changes of the smooth texture of plastic and the wrinkled texture of paper) are used to further highlight the core texture features and make the features of difficult-to-identify targets more recognizable.

[0038] Example 2 This invention provides a high-precision, lightweight classification and detection method for multiple types of waste, comprising the following steps: S1, waste is transported via a conveyor belt of an automated detection structure, and a camera simultaneously acquires waste images and transmits them to a waste target detection model; S2, the waste target detection model first adjusts the image brightness and contrast through an image preprocessing module, and then performs feature extraction and classification recognition on the waste images through the synergistic effect of BiFPN trainable weight allocation and GhostConv feature filtering, outputting classification decision results; S3, the waste sorting structure receives the classification decision results, uses sensor-assisted positioning, and the execution mechanism drives the robotic arm and end effector to complete waste sorting.

[0039] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A high-precision, lightweight classification and detection device for multiple types of waste, characterized in that, include: Waste target detection model, automated detection structure, and waste sorting structure; The garbage target detection model is a lightweight computing platform based on an improvement of YOLOv11. The garbage target detection model integrates GhostConv in the feature extraction stage of YOLOv11 and BiFPN in the feature fusion stage. The automated detection structure is used to transport waste and collect waste images, assisting the lightweight computing platform in classifying and identifying waste; The waste sorting structure receives decision instructions from the waste target detection model and executes sorting actions.

2. The apparatus according to claim 1, characterized in that, The GhostConv filters feature maps of various types of waste by setting feature response thresholds, retaining high-response features corresponding to waste material texture and shape contour, and removing low-response features corresponding to stains and shadows.

3. The apparatus according to claim 1, characterized in that, The dynamic weight allocation strategy of BiFPN is as follows: by using trainable weight parameters, the feature layer corresponding to small target garbage is assigned a higher weight than the feature layer of regular target garbage. The weight parameters are obtained by training multiple types of garbage samples.

4. The apparatus according to claim 1, characterized in that, The training dataset of the garbage target detection model has been cleaned and adapted, and contains at least 20 types of common household waste. The data cleaning includes correcting overlapping labeled samples caused by garbage stacking.

5. The apparatus according to claim 4, characterized in that, The data cleaning process also includes supplementing garbage scene samples with stains, deformation, and damage to obtain a high-quality training dataset.

6. The apparatus according to claim 1, characterized in that, The automated detection structure includes a conveyor belt and a camera. The conveyor belt's transport speed is matched to the end-to-end delay of the waste target detection model, and the camera's installation height and resolution are adapted to image acquisition in the waste stacking scenario.

7. The apparatus according to claim 6, characterized in that, The image acquisition of the camera, which is adapted to the installation height and resolution of the garbage stacking scene, specifically includes: the camera is installed at an angle and shooting downwards, and is configured with an adjustable focal length to adapt to the height range of the garbage stacking, and establishes a connection with the computing platform through a high-speed data interface; The high-speed data interface is configured with a dynamic bandwidth allocation mechanism and a feature packet transmission protocol, which encapsulates the acquired image data according to the upper layer exposed features and the lower layer edge features, and marks the location information respectively. The high-speed data interface maintains real-time interaction with the computing platform. When the computing platform detects that the lower-layer edge feature data is incomplete, it triggers the camera to take a supplementary picture of the target area. After receiving the complete data, the computing platform synchronously performs hierarchical feature pre-labeling and edge enhancement preprocessing operations.

8. The apparatus according to claim 1, characterized in that, The waste sorting structure includes an execution mechanism and sensors. The execution mechanism controls the movement of the robotic arm and end effector through a controller. The sensors are used to assist in the positioning of waste and work together with the automated detection structure and the waste target detection model to form a detection-sorting closed loop.

9. The apparatus according to claim 1, characterized in that, The garbage target detection model is equipped with an image preprocessing module and feature enhancement processing logic; The image preprocessing module uses an illumination compensation algorithm to adjust the brightness and contrast of garbage images collected in dimly lit street environments. The feature enhancement processing logic first enlarges the features of small and difficult-to-identify garbage targets, then strengthens the target edges by adjusting the contrast of feature channels, and finally performs feature highlighting processing based on the grayscale difference of texture features to adapt to the detection requirements of relevant targets.

10. A high-precision, lightweight classification and detection method for multiple types of waste, characterized in that, Using the apparatus according to any one of claims 1-9, the steps include: S1. Waste is transported via a conveyor belt with an automated detection structure, and cameras simultaneously capture images of the waste and transmit them to the waste target detection model. S2. The garbage target detection model first adjusts the brightness and contrast of the image through the image preprocessing module, and then performs feature extraction and classification recognition of the garbage image through the synergistic effect of BiFPN trainable weight allocation and GhostConv feature filtering, and outputs the classification decision result. S3. The waste sorting structure receives the classification decision results, uses sensors for assisted positioning, and drives the robotic arm and end effector to complete the waste sorting.

Images