Waste-oriented detection and classification system

Abstract

The invention discloses a waste-oriented detection and classification system, and particularly relates to the technical field of waste detection, if a sensor module detects waste, a signal is output to trigger an intelligent identification module to work, the intelligent identification module performs preliminary garbage classification through a picture shot by a camera, and the picture is transmitted to a main control core through a serial port; the intelligent recognition module and the main control core train a built-in Rep-YOLOv8s model of the intelligent recognition module through the upper computer to adjust algorithm parameter setting, the sensor module provides weight and volume data for the main control core, after data fusion, the data are compared with parameters set in a database, and an accurate classification result is obtained and fed back to the display module in real time. When the classification result is abnormal or the sensor data exceeds a normal range, giving an alarm prompt; one stepping motor is controlled by the stepping motor module to rotate the garbage can to a designated area, and the other stepping motor is controlled to drive the garbage baffle to achieve precise throwing.

Description

Technical Field

[0001] The present invention relates to the technical field of waste detection, and specifically relates to a waste detection and classification system. Background Art

[0002] Wastes have the characteristics of spatial pollution, acute infection and latent infection. The germs remaining in them are extremely harmful, which can directly endanger human health or cause more serious consequences by polluting soil, water area and atmosphere, etc. The types of wastes are complex and need complex pretreatment to be effectively classified and detected for efficient resource utilization. In this context, the research and development of waste intelligent classification systems is particularly important. Summary of the Invention

[0003] For this reason, the present invention provides a waste detection and classification system, which deeply integrates advanced Internet of Things (IoT) technology and artificial intelligence (AI) technology. By implementing precise classification and real-time monitoring of wastes, it effectively reduces environmental pollution and significantly reduces the potential harmfulness of emissions. The application of IoT technology enables the system to collect, transmit and process waste-related data in real time, providing a rich and accurate data basis for AI algorithms, and then realizing precise identification and scientific classification of wastes to solve the problems raised in the background art.

[0004] To achieve the above object, the present invention provides the following technical solution: A waste detection and classification system, which is composed of an intelligent recognition module, a sensor module, a data processing module, a display module and a stepper motor module;

[0005] If the infrared sensor of the sensor module detects waste, it outputs a signal to trigger the operation of the intelligent recognition module. The intelligent recognition module conducts preliminary waste classification on the pictures taken by the camera and transmits them to the main control core (single-chip microcomputer) of the data processing module through the serial port. The intelligent recognition module and the main control core train the Rep-YOLOv8s waste detection model and adjust the algorithm parameter settings through the upper computer. The weight sensor and ultrasonic proximity sensor of the sensor module respectively provide weight and volume data for the main control core (single-chip microcomputer). After data fusion, it is compared with the parameters set in the database to obtain an accurate classification result; and the classification result, weight and volume information are fed back to the display module in real time. When the classification result is abnormal or the sensor data exceeds the normal range, an alarm prompt can be issued in time; according to the classification result, the main control core controls a stepper motor to rotate the trash can to the designated area through the ULN2003 driver board of the stepper motor module, and controls another stepper motor to drive the trash baffle to achieve precise placement;

[0006] Intelligent recognition module. This module captures waste through an external camera. Considering the characteristics of waste with variable sizes, shapes, and combinations, on the basis of the traditional YOLOv8s object detection algorithm, RepECA is adopted as the backbone network, and the convolutional layers in the original backbone network are replaced with RepVGG modules. At the same time, the ECA attention mechanism and the eSE attention mechanism are incorporated to construct a waste classification model, Rep-YOLOv8s. The Rep-YOLOv8s waste detection model is used to classify waste images;

[0007] The architecture design of the Rep-YOLOv8s waste detection model is as follows:

[0008] Backbone network, integrating the RepVGG module and the eSE module;

[0009] Through structural re-parameterization, the RepVGG module combines the convolution and BN layers into a single-path structure to separate the training model from the inference model and increase the forward inference speed of the network; during inference, it is transformed into a single-path structure similar to VGG;

[0010] The derivation process of the combination of the convolution and BN layers in structural re-parameterization is as follows:

[0011] First, perform convolution. The convolution formula is as follows:

[0012] Conv(x) = w(x) + b (1)

[0013] In the formula: w(x) is the weight function, and b is the bias;

[0014] For the i-th channel of the feature map, the calculation formula of the BN layer is shown in Equation (2):

[0015]

[0016] In the formula: μ i is the mean, σ 2 is the variance, β is the translation factor, γ is the size scaling factor, and ε is a constant; substituting Equation (1) into Equation (2) gives the calculation formula shown in Equation (3), where μ is the mean after substituting Equation (2):

[0017]

[0018] The weight corresponding to the new combined convolutional layer is shown in Equation (4):

[0019]

[0020] The new bias corresponding to the convolutional kernel is shown in Equation (5):

[0021]

[0022] Substituting equation (4) into equation (5) gives:

[0023] BN = W fused + B fused (6)

[0024] where, W fused refers to the weights corresponding to the new merged convolutional layer, and B fused refers to the new bias of the convolutional kernel corresponding to the new merged convolutional layer;

[0025] The ECA attention mechanism is an improved version of the compression and excitation network SE, which reduces the model's parameters while ensuring performance and is more lightweight; based on the given channel dimension C, the adaptive determination method of the convolutional kernel size k is shown in formula (7):

[0026]

[0027] where, |t| odd is the odd number closest to t; γ = 2; b = 1;

[0028] For the convolutional layer, feature extraction is performed using CSPDarknet as the backbone network; this CSPDarknet is divided into two parts, each part containing multiple residual blocks for extracting the basic features of the image; in CSP Darknet, the C3 module is replaced by the C2f module; the C2f module divides the input feature map into two branches for dimensionality reduction processing. The two branches are the direct transmission branch and the v8_C2fBottleneck processing branch. A part of the split feature map is transmitted through the direct transmission branch to the final Concat splicing block without additional processing, while the other part of the split feature map is transmitted to multiple v8_C2fBottleneck blocks for further processing to extract higher-level feature representations; the outputs of the two branches are stacked to form a higher-dimensional feature map and fused through a convolutional layer to obtain more gradient flow information; to extract features of different scales, the Rep-YOLOv8s waste detection model uses the fast spatial pyramid pooling SPPF structure, which effectively reduces the number of model parameters and computational complexity while improving the efficiency of feature extraction;

[0029] In the C2f module of the Rep-YOLOv8s waste detection model, the input feature map is first split by the first convolutional layer to form two parts: one part directly passes through the Bottleneck structure, and the other part is split again after each operation layer to generate skip connections; the diversified feature maps of the direct transmission branch and the v8_C2fBottleneck processing branch finally converge at the eSE module;

[0030] Neck network, using the fast spatial pyramid pooling SPPF structure, and leveraging the network structure that combines FPN and PAN, to effectively fuse waste information between different levels, and then construct a detailed waste feature map;

[0031] Output layer, the predicted waste feature map output by the neck network then enters the prediction stage. In the prediction stage, redundant prediction boxes are removed through a screening mechanism to ensure the accuracy of the waste feature map;

[0032] Sensor module, this module includes an infrared sensor, a weight sensor, and a volume sensor; the infrared sensor is used to detect whether the waste is put into the designated position. If the sensor detects the waste, it outputs a signal for the intelligent recognition module to work; the weight sensor combines multiple elastic bodies that generate deformation when stressed with resistance strain gauges that sense this deformation, so that the weight of the waste is converted into an output electrical signal and actual weight data is generated; the volume sensor measures the volume by calculating the time difference between the transmitted and received echoes, and can reduce costs while maintaining good measurement performance; finally, the volume sensor transmits the converted signal to the data processing module for preprocessing calculations; thus, this device ultimately realizes the function of obtaining waste-related data in real time, providing a basis for further classification and processing;

[0033] Data processing module, with the STM32 system-on-chip as the main control core, the main control core links the three major functional parts and coordinates their stable operation; its operation logic is: the TCRT5000 infrared sensor, the HBMC16 gravity sensor, and the 40Hz ultrasonic proximity sensor are used for detection, and after being parsed by the main control core STM32 system-on-chip, they are sent to the OLED display screen and the mobile phone APP for display; this module is responsible for uniformly collecting the data of the intelligent recognition module and the sensor module, and analyzing and processing them in combination with a dedicated algorithm; when this module receives signals from the sensor module (including weight, ultrasonic proximity sensor) and the intelligent recognition module, it first compares the data with the parameters set in the database, and then verifies again through data fusion to obtain an accurate classification result; immediately afterwards, the data processing module will generate a summary file for this classification and save it in the database, and users can query the classification records to trace the origin of the waste; and as the data continues to accumulate, the classification accuracy will become higher and higher; in other words, this module can help the device learn autonomously and improve the product effect;

[0034] Display module, visually display waste classification, support real-time viewing of classification, weight, and volume information, and have an abnormal alarm function; when the classification result is abnormal or the sensor data exceeds the normal range, it can issue an alarm prompt in a timely manner;

[0035] The stepper motor module is used to control the rotation of the trash can. After the main control core STM32 system-on-chip judges according to the received data, it controls a stepper motor to rotate the trash can to the specified area and controls another stepper motor to drive the trash baffle to achieve precise dumping.

[0036] Preferably, the resolution of the external camera of the intelligent recognition module for photographing waste is 960*640 pixels.

[0037] Preferably, the input layer of the Rep-YOLOv8s waste detection model uses the Mosaic data augmentation method to randomly fuse different images to improve the target recognition ability of the model in complex backgrounds; the letterbox adaptive image scaling technology is introduced to improve the accuracy and stability of target detection.

[0038] Preferably, the weight sensor of the sensor module is the HBMC16 single-point weighing sensor suitable for small machinery; the volume sensor is a 40kHz ultrasonic proximity sensor.

[0039] Preferably, the driving mode of the stepper motor in the stepper motor module is 4-phase 8-beat, the maximum speed is about 14 revolutions per minute, the rated voltage is 12V, and the pull-in torque is not less than 34.3 mN·m. This motor meets the system operation requirements and features fast speed, small power, and low noise.

[0040] Preferably, to quantify the overlap degree between the predicted bounding box and the ground truth box and at the same time solve the problems of slow convergence rate and insufficient positioning accuracy of the traditional loss function, the bounding box regression improved loss function is calculated as shown in formula (8);

[0041] L Enhanced =L CIOU +lambda_1×L Angle +lambda_2×L AspectRatio (8)

[0042] Where L Enhanced is the total loss value; L CIOU is the CIOU loss; L Angle is the angle loss; L AspectRatio is the aspect ratio loss, which is used to maintain the shape stability of the bounding box; lambda_1 and lambda_2 are weight factors.

[0043] The present invention has the following advantages:

[0044] While working on reducing the model complexity, to ensure that the detection and segmentation results are not affected, the backbone network of the Rep-YOLOv8s waste detection model integrates the RepVGG module and the ECA attention mechanism, enabling the network to adaptively adjust the weights of channel features. As a result, in the detection task after training, through structural reparameterization, the original multi-branch structure is transformed into a more concise single-path structure. This not only simplifies the network structure but also effectively avoids unnecessary waste of computing resources in the multi-branch structure. In addition, in the C2f module of the Rep-YOLOv8s waste detection model, the present invention incorporates the eSE self-attention module. This enhancement strategy effectively prevents the loss of channel information caused by the reduction in the number of channels, thereby further improving the model performance and achieving high efficiency in resource utilization in terms of accuracy. Description of the Drawings

[0045] Figure 1 Schematic diagram of the system structure provided in this embodiment;

[0046] Figure 2 Schematic diagram of the intelligent recognition module provided in this embodiment;

[0047] Figure 3 Schematic diagram of the HBMC16 single-point weighing sensor in the sensor module provided in this embodiment;

[0048] Figure 4 Schematic diagram of the 40HZ ultrasonic proximity sensor and the weighing collection box sensor structure in the sensor module provided in this embodiment;

[0049] Figure 5 Schematic diagram of the display module structure provided in this embodiment;

[0050] Figure 6 Schematic diagram of the stepper motor module structure provided in this embodiment;

[0051] Figure 7 Schematic diagram of the Rep-YOLOv8s waste detection model provided in this embodiment;

[0052] Figure 8 Schematic diagram of the Mosaic data augmentation provided in this embodiment;

[0053] Figure 9 Letterbox adaptive image provided in this embodiment;

[0054] Figure 10 Schematic diagram of the waste feature network provided in this embodiment;

[0055] Figure 11 Structure of the RepVGG module provided in this embodiment;

[0056] Figure 12 The working flowchart of the eSE module provided in this embodiment. Detailed implementation manners

[0057] The following specific embodiments illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0058] For example, a medical waste automatic feeding system and its scheduling method with application number CN202210531942.7, a system and method for disposing of medical waste with application number CN201910879856.3, a medical waste crushing and sterilization cabinet with application number CN201910851386.X, and a medical waste collection and safe treatment equipment with application number CN202011080294.5 mainly innovate in the conveying and treatment structures. The medical waste treatment method still uses manual classification and simple landfill, crushing or incineration methods. This method is not only inefficient but also prone to cross-infection and environmental pollution problems.

[0059] Therefore, this embodiment provides a detection and classification system for medical waste. The system consists of an intelligent recognition module, a sensor module, a data processing module, a display module, and a stepping motor module;

[0060] As Figure 1 shown, if the infrared sensor of the sensor module detects medical waste, it outputs a signal to trigger the operation of the intelligent recognition module. The intelligent recognition module performs preliminary garbage classification on the pictures taken by the camera and transmits them to the main control core (single-chip microcomputer) of the data processing module through the serial port. The intelligent recognition module and the main control core train the Rep-YOLOv8s waste detection model through the upper computer and adjust the algorithm parameter settings. The weight sensor and ultrasonic proximity sensor of the sensor module respectively provide weight and volume data for the main control core (single-chip microcomputer). After data fusion, it is compared with the parameters set in the database to obtain an accurate classification result; and the classification result, weight, and volume information are real-time fed back to the display module. When the classification result is abnormal or the sensor data exceeds the normal range, an alarm prompt can be issued in a timely manner; according to the classification result, the main control core controls a stepping motor to rotate the trash can to the specified area through the ULN2003 driver board of the stepping motor module, and controls another stepping motor to drive the trash baffle to achieve precise placement;

[0061] As Figure 2 、 Figures 10 - 11As shown in the figure, first, the label Img tool is used to accurately label the dataset to generate a mask image, which is divided into an 80% training set and a 20% test set. Subsequently, image enhancement techniques such as random rotation and adding noise are applied for data preprocessing and dataset augmentation. The intelligent recognition module captures waste through an external camera with a resolution of 960*640 and inputs the image into the Rep-YOLOv8s waste detection model for classification; this module captures medical waste through an external camera. In view of the characteristics of the variable size, shape, and combination of medical waste, RepECA is used as the backbone network based on the traditional YOLOv8s object detection algorithm, and the convolution layer in the original backbone network is replaced by using the RepVGG module. At the same time, the ECA attention mechanism and the eSE attention mechanism are incorporated to construct the medical waste classification model Rep-YOLOv8s, and the Rep-YOLOv8s waste detection model is used to classify medical waste images;

[0062] As Figure 7 shown, the architecture design of the Rep-YOLOv8s waste detection model is as follows:

[0063] Backbone network, integrating the RepVGG module and the eSE module;

[0064] Through structural reparameterization, the RepVGG module combines the convolution and BN layers into a single-path structure to separate the training model from the inference model and increase the forward inference speed of the network; during inference, it is transformed into a single-path structure similar to VGG;

[0065] The derivation process of the combination of the convolution and BN layers in the structural reparameterization is as follows:

[0066] First, perform convolution, and the convolution formula is as follows:

[0067] Conv(x) = w(x) + b (1)

[0068] In the formula: w(x) is the weight function, and b is the bias;

[0069] For the i-th channel of the feature map, the calculation formula of the BN layer is shown in Equation (2):

[0070]

[0071] In the formula: μ i is the mean, σ 2 is the variance, β is the translation factor, γ is the size scaling factor, and ε is a constant; substituting Equation (1) into Equation (2) gives the calculation formula shown in Equation (3), where μ is the mean after substituting Equation (2):

[0072]

[0073] The weights corresponding to the new convolutional layer after merging are shown in Equation (4):

[0074]

[0075] The new bias corresponding to the convolutional kernel is shown in Equation (5):

[0076]

[0077] Substituting Equation (4) into Equation (5), we get:

[0078] BN = W fused + B fused (6)

[0079] where W fused refers to the weights corresponding to the new convolutional layer after merging, and B fused refers to the new bias corresponding to the convolutional kernel of the new convolutional layer after merging;

[0080] The ECA attention mechanism is an improved version of the squeeze-and-excitation network SE, which reduces the parameters of the model while ensuring performance and is more lightweight; based on the given channel dimension C, the adaptive determination method of the convolutional kernel size k is shown in Equation (7):

[0081]

[0082] where, |t| odd is the odd number closest to t; γ = 2; b = 1;

[0083] For the convolutional layer, feature extraction is carried out using CSPDarknet as the backbone network; this CSPDarknet is divided into two parts, each part containing multiple residual blocks for extracting the basic features of the image; in CSP Darknet, the C3 module is replaced by the C2f module; the C2f module divides the input feature map into two branches for dimensionality reduction processing. The two branches are the direct transmission branch and the v8_C2fBottleneck processing branch. A part of the split feature map is transmitted through the direct transmission branch to the final Concat splicing block without additional processing, while the other part of the split feature map is transmitted to multiple v8_C2fBottleneck blocks for further processing to extract higher-level feature representations; the outputs of the two branches are stacked to form a higher-dimensional feature map and fused through a convolutional layer to obtain more gradient flow information; in order to extract features of different scales, the Rep-YOLOv8s waste detection model uses the fast spatial pyramid pooling SPPF structure, which effectively reduces the number of parameters and computational amount of the model while improving the efficiency of feature extraction;

[0084] In the C2f module of the Rep-YOLOv8s waste detection model, the input feature map is first split by the first convolutional layer, forming two parts: one part directly passes through the Bottleneck structure, and the other part is split after each operation layer to generate skip connections; the diversified feature maps of the direct transmission branch and the v8_C2fBottle neck processing branch finally converge at the eSE module;

[0085] As Figure 10 shown, the neck network uses the fast spatial pyramid pooling SPPF structure and combines the FPN and PAN network structures to effectively fuse waste information between different levels, thereby constructing a detailed waste feature map;

[0086] Output layer, the predicted waste feature map output by the neck network then enters the prediction stage, and redundant prediction boxes are removed through a screening mechanism during the prediction stage to ensure the accuracy of the waste feature map;

[0087] As Figures 3 - 4 shown, the sensor module includes an infrared sensor, a weight sensor, and a volume sensor; the infrared sensor is used to detect whether medical waste is put into the designated position. If the sensor detects medical waste, it outputs a signal for the intelligent recognition module to work; the weight sensor combines multiple elastic bodies that generate deformation under force with strain gauges that sense this deformation, so that the weight of medical waste is converted into an output electrical signal and actual weight data is generated; the volume sensor measures the volume by calculating the time difference between the transmitted and received echoes, and can reduce costs while maintaining good measurement performance; finally, the volume sensor transmits the converted signal to the data processing module for preprocessing calculations; thus, the device finally realizes the function of real-time obtaining waste-related data, providing a basis for further classification and processing;

[0088] The data processing module uses the STM32 system-on-chip as the main control core. The main control core links the three major functional parts and coordinates their stable operation. Its operation logic is as follows: The TCRT5000 infrared sensor, HBMC16 gravity sensor, and 40Hz ultrasonic proximity sensor perform detections. After being parsed by the main control core STM32 system-on-chip, the data is sent to the OLED display screen and the mobile phone APP for display. This module is responsible for uniformly collecting the data of the intelligent recognition module and the sensor module and performing analysis and processing in combination with a dedicated algorithm. When this module receives signals from the sensor module (including weight and ultrasonic proximity sensor) and the intelligent recognition module, it first compares the data with the parameters set in the database, and then verifies it again through data fusion to obtain an accurate classification result. Immediately afterwards, the data processing module will generate a summary file for this classification and save it in the database. Users can query the classification records to trace the medical waste. Moreover, as the data continues to accumulate, the classification accuracy will become higher and higher. In other words, this module can help the device learn autonomously and improve the product effect.

[0089] As Figure 5 shown, the display module visually displays the classification of medical waste, supports real-time viewing of classification, weight, and volume information, and has an abnormal alarm function. When the classification result is abnormal or the sensor data exceeds the normal range, it can issue an alarm prompt in a timely manner.

[0090] In this embodiment, a 0.91-inch OLED display screen is used for displaying the liquid helium data of the magnetic resonance. Compared with the LCD display screen, OLED has the advantages of self-luminescence, no need for a backlight source, high contrast, thin thickness, wide viewing angle, wide operating temperature range, simple structure and manufacturing process, etc., and has now been widely used in various display scenarios. The driving IC of the display screen is SSD1306, which has a built-in boost function, saving circuit space, reducing the volume of the module and lowering the application conditions. The 0.91-inch OLED display screen is composed of 128×32 dot matrices, that is, 128 columns and 32 rows. Among them, every 8 rows are a page. Through programming, the corresponding character data can be displayed in units of pages.

[0091] At the same time, the OLED display screen used in this embodiment communicates with the main control single-chip microcomputer through the IIC protocol. The control logic is simple. Together with the power supply, only 4 pins are needed to achieve driving, greatly saving the peripheral resources of the single-chip microcomputer and the PCB space, and ensuring the stability, security, and speed of data transmission.

[0092] As Figure 6 shown, the stepping motor module is used to control the rotation action of the trash can. After the main control core STM32 system-on-chip judges according to the received data, it controls a stepping motor to rotate the trash can to the specified area, and controls another stepping motor to drive the trash baffle to achieve accurate placement.

[0093] The infrared sensor is used to detect whether medical waste is put into the designated position. If the sensor detects medical waste, it outputs a signal to trigger the intelligent recognition module to work. The intelligent recognition module conducts preliminary garbage classification through the pictures taken by the camera and transmits the classification result to the main control module (single-chip microcomputer) through the serial port. The intelligent recognition module and the main control center can train the model and adjust the algorithm parameters through the upper computer. The weight sensor and the ultrasonic proximity sensor are respectively used for weighing and providing weight and volume data to the main control module (single-chip microcomputer). After data fusion, the data is compared with the parameters set in the database to obtain an accurate classification result, and the classification result, weight, and volume are fed back to the display module in real time. When the classification result is abnormal or the sensor data exceeds the normal range, an alarm prompt can be issued in a timely manner. In addition, according to the classification result, the main control core controls a stepper motor to rotate the trash can to the designated area through the ULN2003 driver board and controls another stepper motor to drive the trash baffle to achieve precise placement.

[0094] As Figure 8 and Figure 9 shown, the input layer of the Rep-YOLOv8s waste detection model adopts the Mosaic data augmentation method to randomly fuse different images to improve the target recognition ability of the model in complex backgrounds; the letterbox adaptive image scaling technology is introduced to improve the accuracy and stability of target detection.

[0095] The weight sensor of the sensor module is the HBMC16 single-point weighing sensor suitable for small machinery; the volume sensor is a 40kHz ultrasonic proximity sensor.

[0096] The drive mode of the stepper motor in the stepper motor module is 4-phase 8-beat, the maximum speed is about 14 revolutions per minute, the rated voltage is 12V, and the pull-in torque is not less than 34.3 mN·m. This motor meets the operation requirements of the system, and its characteristics include high speed, low power, and low noise. After judging according to the received data, the main control core controls a stepper motor to rotate the trash can to the designated area and controls another stepper motor to drive the trash baffle to achieve precise placement.

[0097] To quantify the overlap degree between the predicted bounding box and the ground truth box and solve the problems of slow convergence rate and insufficient positioning accuracy of the traditional loss function, the bounding box regression improved loss function is calculated as shown in formula (8);

[0098] L Enhanced = L CIOU + lambda_1 × L Angle + lambda_2 × L AspectRatio (8)

[0099] where LEnhanced is the overall loss value; L CIOU is the CIOU loss; L Angle is the angle loss; L AspectRatio is the aspect ratio loss, used to maintain the shape stability of the bounding box; lambda_1 and lambda_2 are weight factors.

[0100] Although the present invention has been described in detail with general descriptions and specific embodiments above, on the basis of the present invention, some modifications or improvements can be made, which are obvious to those skilled in the art. Therefore, these modifications or improvements made without departing from the spirit of the present invention all fall within the scope of the present invention claimed.

Claims

1. A waste-oriented detection and classification system, characterized in that: The system consists of an intelligent recognition module, a sensor module, a data processing module, a display module, and a stepper motor module; If the infrared sensor of the sensor module detects waste, it outputs a signal to trigger the operation of the intelligent recognition module. The intelligent recognition module conducts preliminary waste classification on the pictures taken by the camera and transmits them to the main control core of the data processing module through the serial port. The intelligent recognition module and the main control core train the built-in Rep-YOLOv8s waste detection model of the intelligent recognition module and adjust the algorithm parameters through the upper computer. The weight sensor and ultrasonic proximity sensor of the sensor module respectively provide weight and volume data for the main control core. After data fusion, it is compared with the parameters set in the database to obtain an accurate classification result; And the classification result, weight, and volume information are real-time fed back to the display module. When the classification result is abnormal or the sensor data exceeds the normal range, an alarm prompt can be issued in a timely manner; According to the classification result, the main control core controls a stepper motor to rotate the trash can to the specified area through the stepper motor module and controls another stepper motor to drive the trash baffle to achieve precise placement.

2. The waste-oriented detection and classification system according to claim 1, wherein: The intelligent recognition module takes pictures of waste through an external camera. In view of the characteristics of the variable size, shape, and combination of waste, the Rep-YOLOv8s waste detection model is used to classify the waste images; The architecture design of the Rep-YOLOv8s waste detection model is as follows: Backbone network, integrating the RepVGG module and the eSE module; The RepVGG module combines the convolution and BN layers into a single-path structure through structural reparameterization to separate the training model from the inference model and increase the forward inference speed of the network.

3. The waste-oriented detection and classification system according to claim 2, characterized in that: The derivation process of the combination of the convolution and BN layers in the structural reparameterization is as follows: First, perform convolution. The convolution formula is as follows: Conv(x) = w(x) + b (1) Where: w(x) is the weight function and b is the bias; For the i-th channel of the feature map, the calculation formula of the BN layer is shown in Equation (2): where: μ i is the mean value, σ 2 is the variance, β is the translation factor, γ is the size scaling factor, and ε is a constant; substituting Equation (1) into Equation (2) gives the calculation formula as shown in Equation (3), where μ is the mean value after substituting Equation (2): The weight corresponding to the new convolution layer after combination is shown in Equation (4): The new bias of the corresponding convolution kernel is shown in Equation (5): Substitute Equation (4) into Equation (5) to get: BN = W fused + B fused (6) Among them, W fused refers to the weights corresponding to the new convolutional layer after merging, and B fused refers to the new bias of the convolutional kernel corresponding to the new convolutional layer after merging; The ECA attention mechanism is an improved version of the compression and excitation network SE. Based on the given channel dimension C, the adaptive determination method of the convolution kernel size k is shown in Equation (7): where |t| odd is the odd number closest to t; γ = 2; b = 1; Convolution layer, using CSPDarknet as the backbone network for feature extraction; The CSPDarknet is divided into two parts, each part contains multiple residual blocks for extracting the basic features of the image; In CSP Darknet, the C3 module is replaced by the C2f module; The C2f module divides the input feature map into two branches for dimensionality reduction processing. The two branches are the direct transmission branch and the v8_C2fBottleneck processing branch. In the C2f module of the Rep-YOLOv8s waste detection model, the input feature map is first split by the first convolutional layer, forming two parts: one part directly passes through the Bottleneck structure, and the other part is split after each operation layer to generate skip connections. The diversified feature maps of the direct transmission branch and the v8_C2fBottleneck processing branch finally converge at the eSE module.

4. The waste-oriented detection and classification system according to claim 1, characterized in that: The Rep-YOLOv8s waste detection model also includes a neck network that uses the fast spatial pyramid pooling SPPF structure and combines the FPN and PAN network structures to achieve the fusion of waste information between different levels, thereby constructing a predicted waste feature map.

5. The waste-oriented detection and classification system according to claim 1, characterized in that: The Rep-YOLOv8s waste detection model also includes an output layer. The predicted waste feature map output by the neck network then enters the prediction stage, and redundant prediction boxes are removed through a screening mechanism in the output layer to ensure the accuracy of the waste feature map.

6. The waste-oriented detection and classification system according to claim 5, characterized in that: To quantify the overlap degree between the predicted box and the ground truth box and simultaneously solve the problems of slow convergence rate and insufficient localization accuracy of the traditional loss function, the bounding box regression improves the loss function, and the calculation method is shown in formula (8); L Enhanced = L CIOU + lambda_1 × L Angle + lambda_2 × L AspectRatio (8) where L Enhanced is the overall loss value; L CIOU is the CIOU loss; L Angle is the angle loss; L AspectRatio is the aspect ratio loss, used to maintain the shape stability of the bounding box; lambda_1 and lambda_2 are weight factors.

Images