A method and system for automatically sorting recycled and reused resources combined with intelligent identification

Abstract

The present invention relates to an automatic sorting method and system for recycling and reusing resources combined with intelligent identification, comprising the following steps: constructing a multi-layer sorting pipeline, including a coarse sorting pipeline and a fine sorting pipeline; the coarse sorting pipeline acquires garbage data and pre-processes it; based on the pre-processed data, the garbage is roughly classified based on a coarse classification model, and the location of individual items is detected in combination with a target detection algorithm; based on the coarse classification results and the location of the items, a robotic arm is used to sort the items to the corresponding fine sorting pipeline; each fine sorting pipeline performs a secondary classification on the items; each fine sorting pipeline is ultimately connected to an independent conveyor belt, and the materials are directly transported to the corresponding storage area. The weight of the materials is recorded before entering the storage area, and RFID tags are used to add material category, source area, weight, and processing arrangement information to each pile of garbage. The present invention effectively reduces garbage management costs and improves resource recovery efficiency.

Description

Technical Field

[0001] The present invention relates to the field of intelligent sorting, and in particular to a method and system for automatically sorting recycled and reused resources combined with intelligent identification. Background Art

[0002] Traditional manual sorting methods are inefficient, inaccurate in classification, and have high pollution risks, making it difficult to meet the automation, precision, and efficiency requirements of modern waste disposal.

[0003] While existing automated sorting systems have improved waste sorting efficiency to a certain extent, they still face numerous technical challenges. For example, many systems rely solely on a single sensor (such as an RGB camera) or simple rules for classification, resulting in insufficient accuracy in identifying complex waste components and making it difficult to accurately sort waste of various materials. Furthermore, due to the wide variety of waste types, shapes, and sizes, as well as the presence of contaminants on the surfaces of some waste, existing detection algorithms and equipment lack stability and adaptability, making it difficult to cope with complex and changing sorting scenarios. Furthermore, there is a lack of efficient optimization solutions for acquiring location information for waste items and for the coordinated control of sorting lines, and the collaborative efficiency of robotic arm sorting equipment urgently needs to be improved. Summary of the Invention

[0004] In order to solve the above problems, the purpose of the present invention is to provide a method and system for automatic sorting of recycled and reused resources combined with intelligent identification, which can effectively reduce garbage management costs and improve resource recovery efficiency.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for automatically sorting recycled and reused resources combined with intelligent identification, comprising the following steps:

[0007] Build a multi-layer sorting line, including a rough sorting line and a fine sorting line;

[0008] The rough sorting line uses RGB recognition cameras, infrared sensors, and spectrum analyzers to acquire and pre-process garbage data. Based on the pre-processed data, the rough classification model is used to roughly classify the garbage into plastic, glass, metal, and paper. The location of individual items is then detected using a target detection algorithm.

[0009] According to the rough classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines;

[0010] The subdivided sorting lines include plastic sorting lines, metal sorting lines, glass sorting lines and paper sorting lines; each subdivided sorting line performs secondary sorting on the items;

[0011] Each subdivided sorting line is ultimately connected to an independent conveyor belt, and the materials are transported directly to the corresponding storage area. The weight of the materials is recorded before entering the storage area, and RFID tags are used to add material category, source area, weight and processing arrangement information to each pile of garbage.

[0012] Furthermore, garbage data is obtained and pre-processed through RGB recognition camera, infrared sensor and spectrum analyzer, as follows:

[0013] The RGB image is collected by the RGB recognition camera to obtain the color and texture features of the garbage surface and output the RGB image feature F RGB ;

[0014] Through the infrared sensor, the temperature and reflectivity information of the material are obtained to construct the infrared feature vector F IR ;

[0015] Obtain the absorption or reflection spectrum of garbage at different wavelengths through a spectrum analyzer and construct the characteristic vector F of the spectrum Spectral ;

[0016] For RGB image feature F RGB , infrared feature vector F IR and the eigenvector F of the spectrum Spectral Fusion is performed to obtain the final multimodal features based on deep neural networks:

[0017] Construct a deep neural network input layer and reduce the dimensionality of the features of the three modalities through the embedding layer. The formula is as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] Among them, W RGB 、W IR 、W Spectral is the weight coefficient of RGB image, infrared and spectrum; b RGB 、b IR 、b Spectral are the bias matrices of RGB image, infrared and spectrum respectively; Z RGB 、Z IR 、Z Spectral They are the features of RGB image, infrared and spectral modalities after dimension reduction;

[0022] The fusion layer constructs the final multimodal feature F through feature weighted combination final :

[0023] F final =σ(W fuse [Z RGB ,Z IR ,Z Spectral ]+b fuse );

[0024] Among them, W fuse and b fuse is the weight and bias matrix of the fusion layer; σ() is the activation function.

[0025] Furthermore, the RGB image is collected by the RGB recognition camera to obtain the color and texture features of the garbage surface and output the RGB image feature F RGB , as follows:

[0026] Convert the RGB image to HSV color space and extract the garbage color information. The formula is as follows:

[0027] ;

[0028] ;

[0029] ;

[0030] Where R, G, and B are the pixel values of the red, green, and blue channels respectively, H is the hue, S is the saturation, and V is the brightness; the distribution statistics of H, S, and V are histogram features;

[0031] Extracting texture information through gray-level co-occurrence matrix :

[0032] ;

[0033] Where P(i,j|Δx,Δy) represents the probability of the grayscale values i and j appearing between pixels, and Δx and Δy are the horizontal and vertical distances respectively.

[0034] Calculate contrast, energy, and homogeneity:

[0035] ;

[0036] ;

[0037] ;

[0038] Combine the color histogram and texture features to form the RGB image feature vector:

[0039] F RGB={H,S,V,Contrast,Energy,Homogeneity}.

[0040] Furthermore, the garbage is roughly classified into plastic, glass, metal, and paper based on the coarse classification model, and the location of individual items is detected in combination with the target detection algorithm, as follows:

[0041] The coarse classification model adopts a classification model based on deep learning, and according to the final multimodal feature F final , get the probability output P of coarse classification coarse (c|F final ), c represents category c;

[0042] The target detection algorithm uses RGB images as input and combines the multimodal features F final Optimize category information and predict item categories and bounding box position ;

[0043] Combined with the probability output P of coarse classification coarse And detection classification probability P det， Calculate the final classification probability:

[0044] ;

[0045] Among them, α is the fusion weight.

[0046] Furthermore, the deep learning-based classification model includes an input layer, a feature extraction network, an attention module, and an output layer, as follows:

[0047] The input layer inputs the final multimodal feature F final , the feature extraction network takes the input multimodal features F final Through a layer of weighted linear transformation:

[0048] ;

[0049] Where W1 is the weight matrix of the weighted linear transformation; b1 is the bias vector of the weighted linear transformation; Z1 is the output of the weighted linear transformation;

[0050] Enhance nonlinear feature expression through ReLU activation function:

[0051] ;

[0052] Among them, A1 is the output enhanced by the ReLU activation function;

[0053] Through the n-1 hidden layer, further features are extracted:

[0054] ;

[0055] ;

[0056] ;

[0057] Among them, W2,W n are the weight matrices of the first hidden layer and the n-1 hidden layer respectively; b2, b n are the bias vectors of the first hidden layer and the n-1 hidden layer respectively; Z 2， Z n are the outputs of the first hidden layer and the n-1 hidden layer respectively; A2, A n The outputs of the first hidden layer and the n-1 hidden layer are enhanced by the ReLU activation function;

[0058] The Attention module learns the weight distribution W of multimodal features attn Capture important features:

[0059] ;

[0060] Among them, Softmax is the activation function;

[0061] Perform feature weighted vector calculation to obtain enhanced features :

[0062] ;

[0063] The enhanced features By hidden layer:

[0064] ;

[0065] Among them, W attn-hidden 、b attn-hidden are the attention weights and biases; is the hidden layer representation;

[0066] The final attention feature representation and hidden layer representation A n Combined with:

[0067] ;

[0068] Where λ is the binding ratio;

[0069] The output layer converts the final feature F enhanced Mapping to garbage classification space:

[0070] ;

[0071] Among them, W class 、b class is the mapping weight and bias; Z class is the feature after mapping;

[0072] Use the Softmax function to output the probability distribution of each garbage category:

[0073] .

[0074] Furthermore, the target detection algorithm adopts an improved YOLO model, including multi-scale image feature extraction, feature fusion module, bounding box regression module and classification head, as follows:

[0075] Using RGB image I RGB As input, it passes through the backbone network CSPDarknet to extract multi-scale features:

[0076] ;

[0077] in, It is a convolution feature; Backbone is the backbone network;

[0078] Combining multimodal features F final Get the features after further fusion :

[0079]

[0080] Among them, Concat represents the concatenation operation in the channel dimension;

[0081] The bounding box regression module outputs the predicted normalized bounding box position B=(x,y,w,h), (x,y) is the coordinate of the bounding box, w,h is the width and height of the bounding box respectively. The classification head uses the position encoding mechanism to optimize the positioning accuracy of the bounding box and outputs the center coordinates. and width and height :

[0082] ;

[0083] ;

[0084] Among them, t x , t y , t w , t h They are the center coordinates regressed by the bounding box regression network and width and height The offset of ; σ is the activation function; cx 、c y is the predicted grid cell center coordinate offset; p w 、p h is the reference width and height of the corresponding grid unit;

[0085] The final coordinates of the bounding box are:

[0086] ;

[0087] According to the fused features , predict the category probability distribution of each target through the classification head :

[0088] ;

[0089] Among them, W c b c are the weights and biases of the classification head.

[0090] Furthermore, based on the coarse classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines, as follows:

[0091] According to the classification model output of the final classification result c, the target detection model obtains the actual three-dimensional coordinates (Xreal, Yreal, Zreal) of the object location;

[0092] Crawl phase:

[0093] Calculate the grasping point Ppick = (Xreal, Yreal, Zreal) and the safe point Psafe-pick; use inverse kinematics to solve the joint angle and drive the robotic arm to complete the grasping;

[0094] Sorting stage:

[0095] According to the classification result c, the target sorting point Ptarget=(Xtarget, Ytarget, Ztarget) is found, where (Xtarget, Ytarget, Ztarget) is the three-dimensional coordinate of the target sorting point; the path planning moves the item from the grasping point to the target sorting point;

[0096] Release phase:

[0097] At the height of the target point Psafe-targe, the position is lowered and the items are placed on the corresponding assembly line.

[0098] Preferably, each sub-sorting line performs secondary classification on the items, specifically as follows: the plastic sorting line uses an NIR spectrometer to subdivide different categories; the metal sorting line uses an eddy current separator and a magnetic separation device to separate non-ferromagnetic and ferromagnetic metals; the glass sorting line uses an RGB camera to assist in color sorting; the paper sorting line classifies clean paper and contaminated paper.

[0099] A system for automatically sorting recycled and reused resources combined with intelligent identification includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the automatic sorting method for recycling and reused resources combined with intelligent identification as described above.

[0100] A computer storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the steps in the above-mentioned automatic sorting method for recycling and reuse resources combined with intelligent identification.

[0101] The present invention has the following beneficial effects:

[0102] 1. This method uses RGB images, infrared sensors, and a spectrum analyzer to gradually extract features of garbage surface color, texture, temperature, material reflectivity, and spectral characteristics. It then fuses these features using a hybrid strategy involving normalization, weighting, and dimensionality reduction. This ultimately forms a compact feature vector that can be used for subsequent classification and analysis.

[0103] 2. Based on multimodal fusion features, this invention constructs a deep learning coarse classification model to quickly classify garbage into four categories: plastic, glass, metal, and paper. At the same time, it combines target detection algorithms to detect the location of individual items, achieving simultaneous processing of garbage classification and positioning, reducing operational complexity and providing accurate information for subsequent robotic arm operations.

[0104] 3. This invention divides the sorting task into two parts: coarse sorting (classification by broad category) and fine sorting (classification by subcategory), effectively improving sorting efficiency and accuracy. The coarse sorting line quickly completes the broad classification of waste (plastic, glass, metal, paper), reducing the complexity of fine-grained processing. The fine-grained sorting line further refines the processing and improves resource reuse. BRIEF DESCRIPTION OF THE DRAWINGS

[0105] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0106] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0107] refer to Figure 1,In this embodiment, a method for automatic sorting of recycled and reused resources combined with intelligent identification is provided, comprising the following steps;

[0108] Build a multi-layer sorting line, including a rough sorting line and a fine sorting line;

[0109] The rough sorting line uses RGB recognition cameras, infrared sensors, and spectrum analyzers to acquire and pre-process garbage data. Based on the pre-processed data, the rough classification model is used to roughly classify the garbage into plastic, glass, metal, and paper. The location of individual items is then detected using a target detection algorithm.

[0110] According to the rough classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines;

[0111] The subdivided sorting lines include plastic sorting lines, metal sorting lines, glass sorting lines and paper sorting lines; each subdivided sorting line performs secondary sorting on the items;

[0112] Each subdivided sorting line is ultimately connected to an independent conveyor belt, and the materials are transported directly to the corresponding storage area. The weight of the materials is recorded before entering the storage area, and RFID tags are used to add material category, source area, weight and processing arrangement information to each pile of garbage.

[0113] In this embodiment, garbage data is acquired and pre-processed through an RGB recognition camera, an infrared sensor, and a spectrum analyzer, as follows:

[0114] The RGB image is collected by the RGB recognition camera to obtain the color and texture features of the garbage surface and output the RGB image feature F RGB ;

[0115] Through the infrared sensor, the temperature and reflectivity information of the material are obtained to construct the infrared feature vector F IR ;

[0116] Obtain the absorption or reflection spectrum of garbage at different wavelengths through a spectrum analyzer and construct the characteristic vector F of the spectrum Spectral ;

[0117] For RGB image feature F RGB , infrared feature vector F IR and the eigenvector F of the spectrum Spectral Fusion is performed to obtain the final multimodal features based on deep neural networks:

[0118] Construct a deep neural network input layer and reduce the dimensionality of the features of the three modalities through the embedding layer. The formula is as follows:

[0119] ;

[0120] ;

[0121] ;

[0122] Among them, W RGB 、W IR 、W Spectral is the weight coefficient of RGB image, infrared and spectrum; b RGB 、b IR 、b Spectral are the bias matrices of RGB image, infrared and spectrum respectively; Z RGB , Z IR , Z Spectral They are the features of RGB image, infrared and spectral modalities after dimension reduction;

[0123] The fusion layer constructs the final multimodal feature F through feature weighted combination final :

[0124] F final =σ(W fuse [Z RGB ,Z IR ,Z Spectral ]+b fuse );

[0125] Among them, W fuse and b fuse is the weight and bias matrix of the fusion layer; σ() is the activation function.

[0126] In this embodiment, the RGB image is collected by the RGB recognition camera to obtain the color and texture features of the garbage surface, and the RGB image feature F is output. RGB , as follows:

[0127] Convert the RGB image to HSV color space and extract the garbage color information. The formula is as follows:

[0128] ;

[0129] ;

[0130] ;

[0131] Where R, G, and B are the pixel values of the red, green, and blue channels respectively, H is the hue, S is the saturation, and V is the brightness; the distribution statistics of H, S, and V are histogram features;

[0132] Extracting texture information through gray-level co-occurrence matrix (GLCM) :

[0133] ;

[0134] Where P(i,j|Δx,Δy) represents the probability of the grayscale values i and j appearing between pixels, and Δx and Δy are the horizontal and vertical distances respectively.

[0135] Calculate contrast, energy, and homogeneity:

[0136] ;

[0137] ;

[0138] ;

[0139] Combine the color histogram and texture features to form the RGB image feature vector:

[0140] F RGB ={H,S,V,Contrast,Energy,Homogeneity}.

[0141] In this embodiment, garbage is roughly classified into plastic, glass, metal, and paper based on a coarse classification model, and the location of individual items is detected in conjunction with a target detection algorithm, as follows:

[0142] The coarse classification model adopts a classification model based on deep learning, and according to the final multimodal feature F final , get the probability output P of coarse classification coarse (c|F final ), c represents category c;

[0143] The target detection algorithm uses RGB images as input and combines the multimodal features F final Optimize category information and predict item categories and bounding box position ;

[0144] Combined with the probability output P of the coarse classification coarse And detection classification probability P det， Calculate the final classification probability:

[0145] ;

[0146] Among them, α is the fusion weight.

[0147] In this embodiment, the deep learning-based classification model includes an input layer, a feature extraction network, an attention module, and an output layer, as follows:

[0148] The input layer inputs the final multimodal feature F final , the feature extraction network takes the input multimodal features F final Through a layer of weighted linear transformation:

[0149] ;

[0150] Where W1 is the weight matrix of the weighted linear transformation; b1 is the bias vector of the weighted linear transformation; Z1 is the output of the weighted linear transformation;

[0151] Enhance nonlinear feature expression through ReLU activation function:

[0152] ;

[0153] Among them, A1 is the output enhanced by the ReLU activation function;

[0154] Through the n-1 hidden layer, further features are extracted:

[0155] ;

[0156] ;

[0157] ;

[0158] Among them, W2,W n are the weight matrices of the first hidden layer and the n-1 hidden layer respectively; b2, b n are the bias vectors of the first hidden layer and the n-1 hidden layer respectively; Z 2， Z n are the outputs of the first hidden layer and the n-1 hidden layer respectively; A2, A n The outputs of the first hidden layer and the n-1 hidden layer are enhanced by the ReLU activation function;

[0159] The Attention module learns the weight distribution W of multimodal features attn Capture important features:

[0160] ;

[0161] Among them, Softmax is the activation function;

[0162] Perform feature weighted vector calculation to obtain enhanced features :

[0163] ;

[0164] The enhanced features By hidden layer:

[0165] ;

[0166] Among them, W attn-hidden 、b attn-hidden are attention weights and biases; is the hidden layer representation;

[0167] The final attention feature representation and hidden layer representation A n Combined with:

[0168] ;

[0169] Where λ is the binding ratio;

[0170] The output layer converts the final feature F enhanced Mapping to garbage classification space:

[0171] ;

[0172] Among them, W class 、b class is the mapping weight and bias; Z class is the feature after mapping;

[0173] Use the Softmax function to output the probability distribution of each garbage category:

[0174] .

[0175] In this embodiment, the target detection algorithm adopts an improved YOLO model, including multi-scale image feature extraction, feature fusion module, bounding box regression module and classification head, as follows:

[0176] Using RGB image I RGB As input, it passes through the backbone network CSPDarknet to extract multi-scale features:

[0177] ;

[0178] in, It is a convolution feature; Backbone is the backbone network;

[0179] Combining multimodal features F final Get the features after further fusion :

[0180]

[0181] Among them, Concat represents the concatenation operation in the channel dimension;

[0182] The bounding box regression module outputs the predicted normalized bounding box position =(x,y,w,h), (x,y) is the coordinate of the bounding box, w,h are the width and height of the bounding box respectively. The classification head uses the position encoding mechanism to optimize the positioning accuracy of the bounding box and output the center coordinates and width and height :

[0183] ;

[0184] ;

[0185] Among them, t x , t y , t w , t h They are the center coordinates regressed by the bounding box regression network and width and height The offset of ; σ is the activation function; c x 、c y is the predicted grid cell center coordinate offset; p w 、p h is the reference width and height of the corresponding grid unit;

[0186] The final coordinates of the bounding box are:

[0187] ;

[0188] According to the fused features , predict the category probability distribution of each target through the classification head :

[0189] ;

[0190] Among them, W c 、b c are the weights and biases of the classification head.

[0191] In this embodiment, based on the coarse classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines, as follows:

[0192] According to the classification model output of the final classification result c, the target detection model obtains the actual three-dimensional coordinates (Xreal, Yreal, Zreal) of the object location;

[0193] Crawl phase:

[0194] Calculate the grasping point Ppick = (Xreal, Yreal, Zreal) and the safe point Psafe-pick; use inverse kinematics to solve the joint angle and drive the robotic arm to complete the grasping;

[0195] Sorting stage:

[0196] According to the classification result c, the target sorting point Ptarget=(Xtarget, Ytarget, Ztarget) is found, where (Xtarget, Ytarget, Ztarget) is the three-dimensional coordinate of the target sorting point; the path planning moves the item from the grasping point to the target sorting point;

[0197] Release phase:

[0198] At the height of the target point Psafe-targe, the position is lowered and the items are placed on the corresponding assembly line.

[0199] The sub-sorting lines perform secondary classification on items, as follows: the plastic sorting line uses an NIR spectrometer to subdivide different categories; the metal sorting line uses an eddy current separator and a magnetic separation device to separate non-ferromagnetic and ferromagnetic metals; the glass sorting line uses an RGB camera to assist in color sorting; and the paper sorting line classifies clean paper and contaminated paper.

[0200] A system for automatically sorting recycled and reused resources combined with intelligent identification includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps of the automatic sorting method for recycling and reused resources combined with intelligent identification as described above.

[0201] A computer storage medium storing a plurality of instructions adapted to be loaded by a processor and executed by a processor in a method for automatically sorting recycled and reused resources combined with intelligent identification as described above

[0202] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0203] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0204] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0205] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0206] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for automatic sorting of recycled and reused resources combined with intelligent identification, characterized in that: The following steps are included: Build a multi-layer sorting line, including a rough sorting line and a fine sorting line; The rough sorting line uses RGB recognition cameras, infrared sensors, and spectrum analyzers to acquire and pre-process garbage data. Based on the pre-processed data, the rough classification model is used to roughly classify the garbage into plastic, glass, metal, and paper. The location of individual items is then detected using a target detection algorithm. According to the rough classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines; The subdivided sorting lines include plastic sorting lines, metal sorting lines, glass sorting lines and paper sorting lines; each subdivided sorting line performs secondary sorting on the items; Each subdivided sorting line is ultimately connected to an independent conveyor belt, and the materials are transported directly to the corresponding storage area. The weight of the materials is recorded before entering the storage area, and RFID tags are used to add material category, source area, weight and processing arrangement information to each pile of garbage. The garbage data is obtained and pre-processed through the RGB recognition camera, infrared sensor and spectrum analyzer as follows: The RGB image is collected by the RGB recognition camera to obtain the color and texture features of the garbage surface and output the RGB image feature F RGB ; Through the infrared sensor, the temperature and reflectivity information of the material are obtained to construct the infrared feature vector F IR ; Obtain the absorption or reflection spectrum of garbage at different wavelengths through a spectrum analyzer and construct the characteristic vector F of the spectrum Spectral ; For RGB image feature F RGB , infrared feature vector F IR and the eigenvector F of the spectrum Spectral Fusion is performed to obtain the final multimodal features based on deep neural networks: Construct a deep neural network input layer and reduce the dimensionality of the features of the three modalities through the embedding layer. The formula is as follows: ； ； ； Among them, W RGB 、W IR 、W Spectral is the weight coefficient of RGB image, infrared and spectrum; b RGB 、b IR 、b Spectral are the bias matrices of RGB image, infrared and spectrum respectively; Z RGB 、Z IR 、Z Spectral They are the features of RGB image, infrared and spectral modalities after dimension reduction; The fusion layer constructs the final multimodal feature F through feature weighted combination final : F final =σ(W fuse [WITH RGB ,WITH IR ,WITH Spectral ]+b fuse ); Among them, W fuse and b fuse is the weight and bias matrix of the fusion layer; σ() is the activation function.

2. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 1 is characterized in that: The RGB recognition camera collects RGB images, obtains the surface color and texture features of the garbage, and outputs the RGB image feature F RGB , as follows: Convert the RGB image to HSV color space and extract the garbage color information. The formula is as follows: ； ； ； Where R, G, and B are the pixel values of the red, green, and blue channels respectively, H is the hue, S is the saturation, and V is the brightness; the distribution statistics of H, S, and V are histogram features; Extracting texture information through gray-level co-occurrence matrix : ； Where P(i,j|Δx,Δy) represents the probability of the grayscale values i and j appearing between pixels, and Δx and Δy are the horizontal and vertical distances respectively. Calculate contrast, energy, and homogeneity: ； ； ； Combine the color histogram and texture features to form the RGB image feature vector: F RGB ={H,S,V,Contrast,Energy,Homogeneity}。 3. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 1 is characterized in that: The rough classification model is used to roughly classify garbage into plastic, glass, metal, and paper, and the location of individual items is detected using the target detection algorithm, as follows: The coarse classification model adopts a classification model based on deep learning, and according to the final multimodal feature F final , get the probability output P of coarse classification coarse (c|F final ), c represents category c; The target detection algorithm uses RGB images as input and combines the multimodal features F final Optimize category information and predict item categories and bounding box position ; Combined with the probability output P of the coarse classification coarse And detection classification probability P det， Calculate the final classification probability: ； Among them, α is the fusion weight.

4. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 3 is characterized in that: The deep learning-based classification model includes an input layer, a feature extraction network, an attention module, and an output layer, as follows: The input layer inputs the final multimodal feature F final , the feature extraction network takes the input multimodal features F final Through a layer of weighted linear transformation: ； Where W1 is the weight matrix of the weighted linear transformation; b1 is the bias vector of the weighted linear transformation; Z1 is the output of the weighted linear transformation; Enhance nonlinear feature expression through ReLU activation function: ； Among them, A1 is the output enhanced by the ReLU activation function; Through the n-1 hidden layer, further features are extracted: ； ； AND n =ReLU(Z n ); Among them, W2,W n are the weight matrices of the first hidden layer and the n-1 hidden layer respectively; b2, b n are the bias vectors of the first hidden layer and the n-1 hidden layer respectively; Z 2， Z n are the outputs of the first hidden layer and the n-1 hidden layer respectively; A2, A n The outputs of the first hidden layer and the n-1 hidden layer are enhanced by the ReLU activation function; The Attention module learns the weight distribution W of multimodal features attn Capture important features: ； Among them, Softmax is the activation function; Perform feature weighted vector calculation to obtain enhanced features : ； The enhanced features By hidden layer: ； Among them, W attn-hidden 、b attn-hidden are attention weights and biases; is the hidden layer representation; The final attention feature representation and hidden layer representation A n Combined with: ； Where λ is the binding ratio; The output layer converts the final feature F enhanced Mapping to garbage classification space: ； Among them, W class 、b class is the mapping weight and bias; Z class is the feature after mapping; Use the Softmax function to output the probability distribution of each garbage category: 。 5. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 3 is characterized in that: The object detection algorithm uses an improved YOLO model, including multi-scale image feature extraction, feature fusion module, bounding box regression module and classification head, as follows: Using RGB image I RGB As input, it passes through the backbone network CSPDarknet to extract multi-scale features: ； in, It is a convolution feature; Backbone is the backbone network; Combining multimodal features F final Get the features after further fusion : ； Among them, Concat represents the concatenation operation in the channel dimension; The bounding box regression module outputs the predicted normalized bounding box position B=(x,y,w,h), (x,y) is the coordinate of the bounding box, w,h is the width and height of the bounding box respectively. The classification head uses the position encoding mechanism to optimize the positioning accuracy of the bounding box and outputs the center coordinates. and width and height : ； ； Among them, t x , t y , t w , t h They are the center coordinates regressed by the bounding box regression network and width and height The offset of ; σ is the activation function; c x 、c y is the predicted grid cell center coordinate offset; p w 、p h is the reference width and height of the corresponding grid unit; The final coordinates of the bounding box are: ； According to the fused features , predict the category probability distribution of each target through the classification head : ； Among them, W c b c are the weights and biases of the classification head.

6. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 1 is characterized in that: According to the rough classification results and the location of the items, the robotic arm sorts the items to the corresponding fine sorting lines, as follows: According to the classification model output of the final classification result c, the target detection model obtains the actual three-dimensional coordinates (Xreal, Yreal, Zreal) of the object location; Crawl phase: Calculate the grasping point Ppick = (Xreal, Yreal, Zreal) and the safe point Psafe-pick; use inverse kinematics to solve the joint angle and drive the robotic arm to complete the grasping; Sorting stage: According to the classification result c, the target sorting point Ptarget=(Xtarget, Ytarget, Ztarget) is found, where (Xtarget, Ytarget, Ztarget) is the three-dimensional coordinate of the target sorting point; the path planning moves the item from the grasping point to the target sorting point; Release phase: At the height of the target point Psafe-targe, the position is lowered and the items are placed on the corresponding assembly line.

7. The automatic sorting method for recycling and reused resources combined with intelligent identification according to claim 1 is characterized in that: The sub-sorting lines perform secondary classification of items, as follows: the plastic sorting line uses an NIR spectrometer to subdivide different categories; the metal sorting line uses an eddy current separator and a magnetic separation device to separate non-ferromagnetic and ferromagnetic metals; the glass sorting line uses an RGB camera to assist in color sorting; and the paper sorting line classifies clean paper and contaminated paper.

8. An automatic sorting system for recycling and reused resources combined with intelligent identification, characterized in that: The method comprises a processor, a memory and a computer program stored in the memory. When the processor executes the computer program, the method specifically performs the steps of the automatic sorting method for recycling and reuse resources combined with intelligent identification as described in any one of claims 1 to 7.

9. A computer storage medium, characterized in that The computer storage medium stores a plurality of instructions, which are suitable for being loaded by a processor and executing the steps of the automatic sorting method for recycling and reuse resources combined with intelligent identification as described in any one of claims 1 to 7.

Images