Household appliance disassembly part identification method based on convolution attention guidance self-organizing architecture
Through a self-organizing architecture based on convolutional attention guidance, the problem that traditional networks are difficult to capture the characteristics of complex components is solved, efficient and accurate identification of household appliance dismantling parts is achieved, and dismantling efficiency and resource recycling efficiency are improved.
Patent Information
- Application Number
- CN202411186159.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-12-03
AI Technical Summary
Traditional deep learning visual recognition networks are difficult to effectively capture the characteristics of complex and diverse household appliance disassembly parts, resulting in insufficient recognition accuracy and low disassembly efficiency, making it difficult to deal with the problem of accurate identification and classification of massive component data.
Adopting a self-organizing architecture based on convolutional attention guidance, through image acquisition and preprocessing, an efficient self-organizing search strategy for convolutional architecture is designed, combined with dual-optimized self-organizing architecture optimization methods, to build an adaptive network structure and parameters, and optimize it. Model structure to improve identification accuracy.
显著提升了家电拆解零部件的识别准确性和效率,减少了对人工操作的依赖,降低了拆解成本,提高了资源回收率,实现了高效的自动化识别。
Smart Images

Figure CN119068468A_ABST
Abstract
Description
Technical Field
[0001] Based on the image data of household appliance disassembly, the present invention uses a self-organizing architecture guided by convolutional attention to fully learn the features of complex disassembled parts within the class, so that the recognition model can learn the local features of more samples, and optimize the model structure through a self-organizing search strategy to accurately identify household appliance parts with complex styles within the class. In the process of household appliance disassembly, the types and forms of parts are complex and diverse. This self-organizing architecture guided by convolutional attention can effectively learn diverse component feature information, improve the feature capture ability of the model, and achieve high-precision household appliance disassembly component recognition, providing technical support for the disassembly and resource recovery of waste household appliances. This method is an important branch of the field of computer vision, and also belongs to the field of resource recovery and reuse. Background Art
[0002] As the number of waste household appliances increases year by year, the impact on the environment and resources is becoming increasingly significant. Therefore, the recycling and reuse of waste household appliances has become a focus of social concern. As an important part of resource recycling, the dismantling of waste household appliances has become a technical problem that needs to be solved urgently. At present, the dismantling of waste household appliances still mainly relies on manual operation, which is inefficient and costly. In addition, due to the diverse types of waste household appliances and the complex structure of parts, it is difficult to achieve efficient and accurate part classification and identification by manual labor, which limits the development of the waste household appliance recycling industry. Developing an intelligent and automated household appliance disassembly parts identification method is the key to improving the efficiency of waste household appliance recycling. The household appliance disassembly parts identification method based on convolutional neural network effectively improves the level of automated disassembly and reduces manual dependence. However, in the actual household appliance disassembly process, the same type of waste household appliances often contain parts with complex structures and various types. The visual recognition network based on traditional deep learning is difficult to effectively capture the feature information of these complex samples, resulting in insufficient recognition accuracy and low disassembly efficiency.
[0003] As the scale of recycling of used household appliances in my country gradually expands, traditional machine learning algorithms have been unable to cope with the problem of accurate identification and classification of massive parts data. With the rapid development of convolutional neural networks and attention mechanisms in the field of computer vision, as well as the continuous improvement of computing power, the problem of automated identification of household appliance disassembled parts has gradually transformed into a data-driven intelligent identification problem. The data-driven convolutional neural network model can adaptively learn the features of parts, improve the classification performance of the model through multiple trainings, and ultimately achieve efficient and accurate parts identification. In the actual disassembly process, due to the complex and varied styles of parts in the same type of used household appliances, the model faces challenges in capturing these diverse features, and traditional convolutional networks are difficult to deal with effectively, resulting in low recognition accuracy.
[0004] In order to solve the above problems, the present invention designs a method for identifying household appliance disassembled parts based on a convolutional attention-guided self-organizing architecture. First, the image input that the model can learn is obtained through image acquisition and preprocessing of household appliance disassembled parts. Then, an efficient convolutional architecture self-organizing search strategy based on attention guidance is designed to build a model infrastructure that can efficiently rely on data features. Finally, based on the dual-optimization self-organizing architecture optimization method, the optimal network architecture is searched by combining network training with architecture search, which significantly improves the recognition accuracy. Through this method, the model can adaptively adjust the network structure and parameters in different disassembly scenarios, and effectively cope with complex and changeable parts classification tasks.
[0005] In practical applications, the present invention will significantly improve the efficiency and accuracy of household appliance disassembly, reduce reliance on manual operations, reduce disassembly costs, and improve resource recovery rates. By adopting a convolutional attention-guided self-organizing architecture, it can cope with the complex and diverse characteristics of waste household appliance parts, achieve efficient automatic identification, provide technical support for the waste household appliance disassembly and recycling industry, and has broad application prospects. Summary of the invention
[0006] The present invention adopts the following technical solutions and implementation steps:
[0007] 1. A method for identifying household appliance disassembled parts based on a convolutional attention-guided self-organizing architecture, characterized in that the household appliance disassembled parts image acquisition and preprocessing, designing an efficient convolutional architecture self-organizing search strategy based on attention guidance, studying a self-organizing architecture optimization method based on dual optimization, and realizing the identification of household appliance disassembled parts generated by efficient self-organizing architecture, including the following steps:
[0008] (1) Image acquisition and preprocessing of household appliance disassembly parts
[0009] Taking the disassembled parts of home appliances as the research object, the disassembled parts of home appliances with complex backgrounds are selected as input images, which contain 13 types of home appliance disassembled parts to be identified from the disassembled parts of the assembly line, and the labels are L = (l1, l2, …, l 13 ), l1 is the condenser, l2 is the capacitor, l3 is the transformer, l4 is the external motor, l5 is the circuit board, l6 is the compressor, l7 is the copper tube, l8 is the wire, l9 is the aluminum, l 10 For drainage pipe, l 11 For flame retardant, 12 For waste, 13 For the indoor fan;
[0010] Original home appliance disassembly parts image data I O , disassemble the image data of the parts O Image data constructed as RGB color space:
[0011] I RGB =[I R ,I G ,I B ] (1)
[0012] Among them, I RGB is the RGB image matrix, I R Represents the red channel pixel matrix, I G Represents the green channel pixel matrix, I B Represents the blue channel pixel matrix;
[0013] The normalization calculation formula of the RGB image matrix is as follows:
[0014]
[0015] Among them, X RGB is the RGB image matrix after normalizing each element of the three-channel matrix; is the maximum pixel value of the RGB image matrix; is the minimum pixel value of the RGB image matrix;
[0016] (2) Design an efficient self-organizing search strategy for convolutional architectures based on attention guidance
[0017] Construct a self-organizing architecture search strategy based on convolutional neural networks, where the search of the self-organizing architecture recognition model includes the design of the search space O, establishes an attention-guided multi-convolution candidate operation mask sharing architecture search, and constructs a channel partial link architecture search method;
[0018] Design of search space O: The search space consists of 9 candidate operations, O = (o1, o2, ... o9); o1 is a 1×1 convolution operation, o2 is a 3×3 convolution operation, o3 is a 5×5 convolution operation, o4 is a 3×3 maximum pooling operation, o5 is a 3×3 average pooling operation, o6 is a 3×3 dilated convolution operation, o7 is a 5×5 dilated convolution operation, o8 is a skip connection operation, and o9 is an identity zeroing operation;
[0019] Establish an attention-guided multi-convolution candidate operation mask sharing architecture search method: The self-organizing architecture based on the convolutional neural network is composed of N nodes, the starting node of the self-organizing architecture is i, the ending node of the self-organizing architecture is j, i = 0, 1, ..., N, j = 1, 2, ..., N, N = 8 and i < j, the multi-convolution candidate operation mask sharing calculation formula for the i-th to j-th nodes is as follows:
[0020]
[0021] in, represents the input feature of the i-th node at the t-th iteration; represents the output feature of the jth node at the tth iteration; when i = 1, the input is the input image matrix X at the tth iteration RGB (t); f(·) represents the composite function of the operation from the i-th node to the j-th node; o m Represents the operation between the i-th node and the j-th node in the search space O; represents the weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration; Represents the nth convolution candidate operation mask at the tth iteration; represents the attention weight of the nth convolution candidate operation mask at the tth iteration;
[0022] The calculation formula for mapping the weight of the operation between the i-th node and the j-th node to [0,1] is as follows:
[0023]
[0024] in, represents the normalized weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration;
[0025] The calculation formula of the attention weight of the nth convolution candidate operation mask is as follows:
[0026]
[0027] in, Indicates the first layer of attention learning weights at the tth iteration; Indicates the second layer attention learning weight at the tth iteration; δ represents the ReLU activation function; softmax(·) represents Convert it into a probability distribution where each element has a value range between (0,1) and the sum of all elements is 1;
[0028] Construct channel partial link architecture search method: Use partial channel mask to construct the candidate operation architecture search of channel partial link. The candidate convolution search strategy calculation formula for nodes i to j is as follows:
[0029]
[0030] in, Indicates the partial channel mask from the i-th node to the j-th node at the t-th iteration;
[0031] The partial channel mask is calculated as follows:
[0032]
[0033] in, Indicates the calculation operation of randomly selecting a quarter of the number of feature channels; 1 and 0 respectively indicate that at the tth iteration, some channel masks The mask value of
[0034] (3) Research on the optimization method of self-organizing architecture based on dual optimization
[0035] The self-organizing architecture operation based on the convolutional neural network performs dual optimization by alternating between architecture search and architecture training. The calculation formula for optimizing the candidate operation is as follows:
[0036]
[0037] in, represents the training loss function of the self-organizing architecture at the tth iteration; They represent the verification loss function of the self-organizing architecture at the tth iteration; ω t represents the weight of the operation parameters in the architecture training phase at the tth iteration; represents the weight of the operation parameters in the architecture verification phase at the tth iteration;
[0038] The calculation formula for selecting the candidate convolution operation for nodes i to j is as follows:
[0039]
[0040] in, represents the final candidate operation selected by the self-organizing architecture according to the maximum weight at the tth iteration;
[0041] (4) Identification of household appliance disassembly parts
[0042] The specific process of identifying household appliance disassembly parts using the efficient self-organizing architecture based on convolutional neural networks is as follows:
[0043] 1) Use the collected images of disassembled home appliance parts as training data and select the image matrix I O , by normalizing the RGB three-channel pixel matrix, we get the input data X for the tth iteration RGB (t);
[0044] 2) Through the architecture search process shown in formulas (3) to (10), a home appliance disassembly parts recognition model based on convolutional attention guidance is self-organized;
[0045] 3) The images of the disassembled home appliance parts are input into the self-organized home appliance disassembled parts recognition model to obtain the classified label value L of the disassembled home appliance parts, which is the category of the disassembled home appliance parts.
[0046] The creativity of the present invention is mainly reflected in:
[0047] (1) Aiming at the actual household appliance disassembled parts identification process, the present invention proposes an attention-guided efficient convolutional architecture self-organizing search strategy, which allows the network to use the constituent modules of the attention self-organizing distribution network architecture through the input disassembled parts image data, so that the generated model architecture can better match the parts data disassembled by different factories, effectively improving the network's recognition accuracy of the disassembled parts.
[0048] (2) The present invention proposes a method for identifying household appliance disassembled parts based on a convolutional attention guided self-organizing architecture, which can perform high-precision classification of various types of household appliance disassembled parts through a self-organizing network architecture, thereby realizing intelligent and automated parts classification during the household appliance disassembly process, improving the recognition accuracy and recycling efficiency of household appliance disassembled parts, and at the same time reducing labor costs and improving the resource recycling benefits of the entire society. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 This is a sub-architecture diagram of the home appliance disassembly parts identification model self-organized and constructed by the present invention DETAILED DESCRIPTION
[0050] The present invention adopts the following technical solutions and implementation steps:
[0051] A method for identifying household appliance disassembled parts based on a convolutional attention-guided self-organizing architecture is characterized by collecting and preprocessing images of household appliance disassembled parts, designing an efficient convolutional architecture self-organizing search strategy based on attention guidance, studying a self-organizing architecture optimization method based on dual optimization, and realizing the identification of household appliance disassembled parts generated by efficient self-organizing architecture, including the following steps:
[0052] (1) Image acquisition and preprocessing of household appliance disassembly parts
[0053] Taking the disassembled parts of home appliances as the research object, the disassembled parts of home appliances with complex backgrounds are selected as input images, which contain 13 types of home appliance disassembled parts to be identified from the disassembled parts of the assembly line, and the labels are L = (l1, l2, …, l 13 ), l1 is the condenser, l2 is the capacitor, l3 is the transformer, l4 is the external motor, l5 is the circuit board, l6 is the compressor, l7 is the copper tube, l8 is the wire, l9 is the aluminum, l 10 For drainage pipe, l 11 For flame retardant, 12 For waste, 13 For the indoor fan;
[0054] Original home appliance disassembly parts image data I O , disassemble the image data of the parts O Image data constructed as RGB color space:
[0055] I RGB =[I R ,I G ,I B ] (1)
[0056] Among them, I RGB is the RGB image matrix, I R Represents the red channel pixel matrix, I G Represents the green channel pixel matrix, I B Represents the blue channel pixel matrix;
[0057] The normalization calculation formula of the RGB image matrix is as follows:
[0058]
[0059] Among them, X RGB is the RGB image matrix after normalizing each element of the three-channel matrix; is the maximum pixel value of the RGB image matrix; is the minimum pixel value of the RGB image matrix;
[0060] (2) Design an efficient self-organizing search strategy for convolutional architectures based on attention guidance
[0061] Construct a self-organizing architecture search strategy based on convolutional neural networks, where the search of the self-organizing architecture recognition model includes the design of the search space O, establishes an attention-guided multi-convolution candidate operation mask sharing architecture search, and constructs a channel partial link architecture search method;
[0062] Design of search space O: The search space consists of 9 candidate operations, O = (o1, o2, ... o9); o1 is a 1×1 convolution operation, o2 is a 3×3 convolution operation, o3 is a 5×5 convolution operation, o4 is a 3×3 maximum pooling operation, o5 is a 3×3 average pooling operation, o6 is a 3×3 dilated convolution operation, o7 is a 5×5 dilated convolution operation, o8 is a skip connection operation, and o9 is an identity zeroing operation;
[0063] Establish an attention-guided multi-convolution candidate operation mask sharing architecture search method: The self-organizing architecture based on the convolutional neural network is composed of N nodes, the starting node of the self-organizing architecture is i, the ending node of the self-organizing architecture is j, i = 0, 1, ..., N, j = 1, 2, ..., N, N = 8 and i < j, the multi-convolution candidate operation mask sharing calculation formula for the i-th to j-th nodes is as follows:
[0064]
[0065] in, represents the input feature of the i-th node at the t-th iteration; represents the output feature of the jth node at the tth iteration; when i = 1, the input is the input image matrix X at the tth iteration RGB (t); f(·) represents the composite function of the operation from the i-th node to the j-th node; o m Represents the operation between the i-th node and the j-th node in the search space O; represents the weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration; Represents the nth convolution candidate operation mask at the tth iteration; represents the attention weight of the nth convolution candidate operation mask at the tth iteration;
[0066] The calculation formula for mapping the weight of the operation between the i-th node and the j-th node to [0,1] is as follows:
[0067]
[0068] in, represents the normalized weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration;
[0069] The calculation formula of the attention weight of the nth convolution candidate operation mask is as follows:
[0070]
[0071] in, Indicates the first layer of attention learning weights at the tth iteration; Indicates the second layer attention learning weight at the tth iteration; δ represents the ReLU activation function; softmax(·) represents Convert it into a probability distribution where each element has a value range between (0,1) and the sum of all elements is 1;
[0072] Construct channel partial link architecture search method: Use partial channel mask to construct the candidate operation architecture search of channel partial link. The candidate convolution search strategy calculation formula for nodes i to j is as follows:
[0073]
[0074] in, Indicates the partial channel mask from the i-th node to the j-th node at the t-th iteration;
[0075] The partial channel mask is calculated as follows:
[0076]
[0077] in, Indicates the calculation operation of randomly selecting a quarter of the number of feature channels; 1 and 0 respectively indicate that at the tth iteration, some channel masks The mask value of
[0078] (3) Research on the optimization method of self-organizing architecture based on dual optimization
[0079] The self-organizing architecture operation based on the convolutional neural network performs dual optimization by alternating between architecture search and architecture training. The calculation formula for optimizing the candidate operation is as follows:
[0080]
[0081] in, represents the training loss function of the self-organizing architecture at the tth iteration; They represent the verification loss function of the self-organizing architecture at the tth iteration; ω t represents the weight of the operation parameters in the architecture training phase at the tth iteration; represents the weight of the operation parameters in the architecture verification phase at the tth iteration;
[0082] The calculation formula for selecting the candidate convolution operation for nodes i to j is as follows:
[0083]
[0084] in, represents the final candidate operation selected by the self-organizing architecture according to the maximum weight at the tth iteration;
[0085] (4) Identification of household appliance disassembly parts
[0086] The specific process of identifying household appliance disassembly parts using the efficient self-organizing architecture based on convolutional neural networks is as follows:
[0087] 1) Use the collected images of disassembled home appliance parts as training data and select the image matrix I O , by normalizing the RGB three-channel pixel matrix, we get the input data X for the tth iteration RGB (t);
[0088] 2) Through the architecture search process shown in formulas (3) to (10), a home appliance disassembly parts recognition model based on convolutional attention guidance is self-organized;
[0089] 3) The images of the disassembled home appliance parts are input into the self-organized home appliance disassembled parts recognition model to obtain the classified label value L of the disassembled home appliance parts, which is the category of the disassembled home appliance parts.
Claims
1. A method for identifying household appliance disassembly parts based on a convolutional attention guided self-organizing architecture, characterized in that: The following steps are involved: (1) Image acquisition and preprocessing of household appliance disassembly parts Taking the disassembled parts of home appliances as the research object, the disassembled parts of home appliances with complex backgrounds are selected as input images, which contain 13 types of home appliance disassembled parts to be identified from the disassembled parts of the assembly line, and the labels are L = (l1, l2, …, l 13 ), l1 is the condenser, l2 is the capacitor, l3 is the transformer, l4 is the external motor, l5 is the circuit board, l6 is the compressor, l7 is the copper tube, l8 is the wire, l9 is the aluminum, l 10 For drainage pipe, l 11 For flame retardant, 12 For waste, 13 For the indoor fan; Original home appliance disassembly parts image data I O , disassemble the image data of the parts O Image data constructed in RGB color space: I RGB =[I R ,I G ,I B ](1) Among them, I RGB is the RGB image matrix, I R Represents the red channel pixel matrix, I G Represents the green channel pixel matrix, I B Represents the blue channel pixel matrix; The normalization calculation formula of the RGB image matrix is as follows: Among them, X RGB is the RGB image matrix after normalizing each element of the three-channel matrix; is the maximum pixel value of the RGB image matrix; is the minimum pixel value of the RGB image matrix; (2) Design an efficient self-organizing search strategy for convolutional architectures based on attention guidance Construct a self-organizing architecture search strategy based on convolutional neural networks, where the search of the self-organizing architecture recognition model includes the design of the search space O, establishes an attention-guided multi-convolution candidate operation mask sharing architecture search, and constructs a channel partial link architecture search method; Design of search space O: The search space consists of 9 candidate operations, O = (o1, o2, ... o9); o1 is a 1×1 convolution operation, o2 is a 3×3 convolution operation, o3 is a 5×5 convolution operation, o4 is a 3×3 maximum pooling operation, o5 is a 3×3 average pooling operation, o6 is a 3×3 dilated convolution operation, o7 is a 5×5 dilated convolution operation, o8 is a skip connection operation, and o9 is an identity zeroing operation; Establish an attention-guided multi-convolution candidate operation mask sharing architecture search method: The self-organizing architecture based on the convolutional neural network is composed of N nodes, the starting node of the self-organizing architecture is i, the ending node of the self-organizing architecture is j, i = 0, 1, ..., N, j = 1, 2, ..., N, N = 8 and i < j, the multi-convolution candidate operation mask sharing calculation formula for the i-th to j-th nodes is as follows: in, represents the input feature of the i-th node at the t-th iteration; represents the output feature of the jth node at the tth iteration; when i = 1, the input is the input image matrix X at the tth iteration RGB (t); f(·) represents the composite function of the operation from the i-th node to the j-th node; o m Represents the operation between the i-th node and the j-th node in the search space O; represents the weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration; Represents the nth convolution candidate operation mask at the tth iteration; represents the attention weight of the nth convolution candidate operation mask at the tth iteration; The calculation formula for mapping the weight of the operation between the i-th node and the j-th node to [0,1] is as follows: in, represents the normalized weight of the operations contained in the search space between the i-th node and the j-th node at the t-th iteration; The calculation formula of the attention weight of the nth convolution candidate operation mask is as follows: in, Indicates the first layer of attention learning weights at the tth iteration; Indicates the second layer attention learning weight at the tth iteration; δ represents the ReLU activation function; softmax(·) represents Convert it into a probability distribution where each element has a value range between (0,1) and the sum of all elements is 1; Construct channel partial link architecture search method: Use partial channel mask to construct the candidate operation architecture search of channel partial link. The candidate convolution search strategy calculation formula for nodes i to j is as follows: in, Indicates the partial channel mask from the i-th node to the j-th node at the t-th iteration; The partial channel mask is calculated as follows: in, Indicates the calculation operation of randomly selecting a quarter of the number of feature channels; 1 and 0 respectively indicate that at the tth iteration, some channel masks The mask value of (3) Research on the optimization method of self-organizing architecture based on dual optimization The self-organizing architecture operation based on the convolutional neural network performs dual optimization by alternating between architecture search and architecture training. The calculation formula for optimizing the candidate operation is as follows: in, represents the training loss function of the self-organizing architecture at the tth iteration; They represent the verification loss function of the self-organizing architecture at the tth iteration; ω t represents the weight of the operation parameters in the architecture training phase at the tth iteration; represents the weight of the operation parameters in the architecture verification phase at the tth iteration; The calculation formula for selecting the candidate convolution operation for nodes i to j is as follows: in, represents the final candidate operation selected by the self-organizing architecture according to the maximum weight at the tth iteration; (4) Identification of household appliance disassembly parts The specific process of identifying household appliance disassembly parts using the efficient self-organizing architecture based on convolutional neural networks is as follows: 1) Use the collected images of disassembled home appliance parts as training data and select the image matrix I O , by normalizing the RGB three-channel pixel matrix, we get the input data X for the tth iteration RGB (t); 2) Through the architecture search process shown in formulas (3) to (10), a home appliance disassembly parts recognition model based on convolutional attention guidance is self-organized; 3) The images of the disassembled home appliance parts are input into the self-organized home appliance disassembled parts recognition model to obtain the classified label value L of the disassembled home appliance parts, which is the category of the disassembled home appliance parts.
