Waste household appliance disassembly part identification method based on sparse representation precision quantification

Abstract

The invention provides a waste household appliance disassembly part identification method based on sparse representation precision quantification, aiming at the problem that decision features and weight distribution are unstable in the quantification process of a waste household appliance disassembly part identification model. The method comprises the following steps: firstly, constructing a waste household electrical appliance disassembly part identification model by using ConvNeXt-Tiny, calculating to obtain depth features of a waste household electrical appliance image, secondly, optimizing a weight in a quantitative training process by using a quantitative error loss function, and finally, freezing an unstable bit width of the model in the training process by using a distribution loss function so as to realize accurate classification of waste household electrical appliances. And the identification precision of the model is improved. According to the waste household appliance disassembly part identification method based on sparse representation precision quantification, in the actual waste household appliance disassembly process, the problem that decision features and weight distribution of an identification model are unstable in the quantification process can be solved, efficient classification and identification of waste household appliance disassembly parts are achieved, and technical support is provided for the waste household appliance recovery industry.

Description

Technical Field

[0001] Based on sparse representation precision quantization, the present invention proposes a method for identifying disassembled parts of waste household appliances. This method collects data of disassembled parts of waste household appliances, constructs an identification model based on ConvNeXt-Tiny, and uses sparse representation precision quantization to train the model, realizing efficient classification and identification of disassembled parts of waste household appliances. This technology solves the problem of unstable decision-making features and weight allocation of the identification model caused by the quantization process during the disassembly process of waste household appliances, providing reliable technical support for the resource utilization and environmental protection of waste household appliances. Background Art

[0002] With the rapid development of technology and the replacement of household appliances, the replacement speed of household appliances has increased significantly. Disassembled parts of waste household appliances contain a large amount of recyclable resources such as metals and plastics, and also contain some harmful substances such as lead, mercury, and cadmium. The correct classification and disassembly of waste household appliances are of great significance for resource recycling. However, there are a wide variety of waste household appliances with different shapes, and manual classification is time-consuming and laborious, and the accuracy cannot be guaranteed. It is urgent to study a method for automatically and accurately identifying the categories of waste household appliances.

[0003] Full-precision deep neural networks have been used for identification in scenes. Using full-precision deep neural networks, accurate identification in the visual domain can be achieved. Full-precision deep neural networks can be used for waste household appliance identification, and the identification performance is better than that of shallow networks. However, full-precision deep neural networks consume a large amount of resources, have poor real-time identification performance, and are difficult to deploy during the waste household appliance identification process. Quantization methods can reduce model complexity and promote the real-time performance of model identification. In the actual disassembly process of waste household appliances, the quantization process causes unstable decision-making features and weight allocation of the identification model, affecting the identification effect.

[0004] The present invention designs a method for identifying disassembled parts of waste household appliances based on sparse representation precision quantization, realizing intelligent and efficient identification of disassembled parts of waste household appliances. This method designs a quantization error loss function to optimize the weights during the quantization training process; secondly, combined with the distribution loss function, the unstable bit widths of the model are frozen during the training process to achieve accurate classification of waste household appliances. The problem of unstable decision-making features and weight allocation of the identification model during the quantization process is solved. Summary of the Invention

[0005] The present invention obtains a method for identifying disassembled parts of waste household appliances based on sparse representation accuracy quantization. This method designs a quantization error loss function to optimize the weights during the quantization training process. Secondly, combined with the distribution loss function, the unstable bit widths of the model are frozen during the training process. This method solves the problem of unstable decision features and weight allocation of the recognition model caused by the quantization process in the actual disassembly process of waste household appliances, and realizes the efficient identification of the types of disassembled parts of waste household appliances. It reduces the cost of identification and better serves the recycling and resource utilization of waste household appliances disassembly.

[0006] The present invention adopts the following technical solutions:

[0007] A method for identifying disassembled parts of waste household appliances based on sparse representation accuracy quantization, characterized by collecting data of disassembled parts of waste household appliances, constructing an identification model for disassembled parts of waste household appliances, training an identification model for disassembled parts of waste household appliances based on sparse representation accuracy quantization, and identifying the types of disassembled parts of waste household appliances, including the following steps:

[0008] (1) Collect data of disassembled parts of waste household appliances

[0009] Obtain an image data set of disassembled parts of waste household appliances, including 12 types of images of disassembled parts of waste household appliances: circuit boards, wires, aluminum, compressors, condensers, transformers, copper pipes, outer motors, drain pipes, flame retardants, waste debris, and inner machine fans, with a total of s JPG-format images. To meet the input requirements of the model, the pixels of each image in the image data set of disassembled parts of waste household appliances are normalized to [0,1];

[0010] Divide the normalized image data set of disassembled parts of waste household appliances into a training set and a test set. The training set is X s , and the test set is 30% of the total number of images is s, and the number of image categories of disassembled parts of waste household appliances in s X is 12;

[0011] (2) Construct an identification model for disassembled parts of waste household appliances

[0012] Construct a waste household appliance disassembly component network based on ConvNeXt-Tiny: The network consists of five parts. The first part, conv1, is composed of a convolutional layer with a kernel size of 1×1, a batch normalization layer, and a GeLU activation function layer, with 96 channels. The second part, conv2, consists of two inverted bottleneck modules. Each inverted bottleneck module is composed of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 96 channels. The third part, conv3, consists of two inverted bottleneck modules. Each inverted bottleneck module is composed of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 192 channels. The fourth part, conv4, consists of two inverted bottleneck modules. Each inverted bottleneck module is composed of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 384 channels. The fifth part, conv5, consists of two inverted bottleneck modules. Each inverted bottleneck module is composed of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 768 channels; X s Define F as the input of the waste household appliance disassembly component network based on ConvNeXt-T θ (X s ; θ t ) as the feature extractor of the waste household appliance disassembly component network. t is the number of iterative training times. θ t is the learnable parameter of F θ (X s ; θ t ). The size of the parameter matrix is 512×28×28. The predictor G ω (F; ω t ) is the output of the waste household appliance disassembly component network. F is the input feature of G ω (F; ω t ). ω t is the weight of G ω (F; ω t ). The size of the weight matrix is 2048×10;

[0013] (3) Train the waste household appliance disassembly component recognition model based on sparse representation accuracy quantization

[0014] Train the waste household appliance disassembly component recognition model based on sparse representation accuracy quantization. Input the training set X s , and after passing through F θ (X s ; θ t ) and G ω (F; ω t)Complete the training of the recognition model, G ω (F; ω t )The expression of the total training loss function L D is:

[0015] L D = L ce + L qe + L bc (1)

[0016] Among them, L qe is the quantization error loss function, L bc is the distribution loss function, L ce is the cross-entropy loss function, and the formula is:

[0017]

[0018] Among them, is the true label of the i-th image data of the j-th class of images, j = 1, 2,..., 12, i = 1, 2,..., N, and N is the number of images of the j-th class, is the predicted probability that image i belongs to class j, Y is the true label of the input image class, P is the predicted probability of the input image class, L qe Calculates the loss of the model parameters before and after quantization, and the formula is:

[0019]

[0020] Among them, t = 1, 2,..., T, and T is the number of training times of the waste household appliance disassembly part recognition model, is the quantized weight, and the corresponding formula is:

[0021]

[0022] Among them, round(·) represents taking an integer, Δ w is the quantization step size, and the formula for Δ w is:

[0023]

[0024] Among them, max(·) represents taking the maximum value, min(·) represents taking the minimum value, and the distribution loss L bc is used to ensure that the output distribution of the quantized model is consistent with that of the unquantized model, and the formula is:

[0025]

[0026] Among them, E is the calculation of the expected value, ||·|| is the norm of the vector, O S is the output of the layer corresponding to the minimum bit width, O His the output of the layer corresponding to the maximum bit width, is the weight sparse representation, and the corresponding formula is:

[0027]

[0028] where ⊙ is channel multiplication, and H(·) is the unit step function, and the formula is:

[0029]

[0030] Use the gradient descent algorithm to optimize the parameters θ t and ω t in the sparse representation accuracy quantization of the waste household appliance disassembly part network, and the parameter update formula is:

[0031]

[0032] where ω t+1 is the parameter matrix of G ω (F; ω t ) during the (t + 1)-th iteration training, and θ t+1 is the parameter matrix of F θ (X t ; θ t ) during the (t + 1)-th iteration training, denotes taking the partial derivative. When t is 300 or more, terminate the training of the recognition model, save the weight parameters of the last training, and complete the training of the recognition model;

[0033] (4) Identify the types of waste household appliance disassembly parts

[0034] After completing the model training, load the weights of the waste household appliance disassembly part recognition model based on sparse representation accuracy quantization, and recognize the images of 12 types of waste household appliance disassembly parts in the test set to obtain the predicted class labels of the waste household appliance disassembly parts and complete the waste household appliance disassembly part recognition task. Description of the Drawings

[0035] Figure 1 Schematic diagram of the error distribution of the waste household appliance disassembly part recognition model with sparse representation accuracy quantization Detailed Embodiments

[0036] A method for identifying waste household appliance disassembly parts based on sparse representation accuracy quantization, characterized by comprising the following steps:

[0037] (1) Collect data of waste household appliance disassembly parts

[0038] Obtain the image dataset of waste household appliance disassembly parts, including 12 types of waste household appliance disassembly part images: circuit boards, wires, aluminum, compressors, condensers, transformers, copper pipes, outer motors, drain pipes, flame retardants, waste debris, and inner machine fans. There are a total of s JPG format images. To meet the input requirements of the model, the pixels of each image in the waste household appliance disassembly part image dataset are normalized to [0, 1];

[0039] Divide the normalized waste household appliance disassembly part image dataset into a training set and a test set. The training set is X s , and the test set is 30% of the total number of images s, and the number of waste household appliance disassembly part image categories in X s is also 12;

[0040] (2) Construct a waste household appliance disassembly part recognition model

[0041] Construct a waste household appliance disassembly part network based on ConvNeXt-Tiny: This network consists of five parts. The first part, conv1, consists of a convolutional layer with a kernel size of 1×1, a batch normalization layer, and a GeLU activation function layer, with 96 channels. The second part, conv2, consists of two inverted bottleneck modules. Each inverted bottleneck module consists of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 96 channels. The third part, conv3, consists of two inverted bottleneck modules. Each inverted bottleneck module consists of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 192 channels. The fourth part, conv4, consists of two inverted bottleneck modules. Each inverted bottleneck module consists of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 384 channels. The fifth part, conv5, consists of two inverted bottleneck modules. Each inverted bottleneck module consists of a depthwise separable convolutional layer with a kernel size of 1×1, a GeLU activation function layer, and a convolutional layer with a kernel size of 1×1, with 768 channels; X s For the input of the waste household appliance disassembly part network based on ConvNeXt-T, define F θ (X s ; θ t ) as the feature extractor of the waste household appliance disassembly part network, where t is the number of iterative training times, and θ t is the learnable parameter of F θ (X s ; θ t ), with the parameter matrix size of 512×28×28, and the predictor Gω (F; ω t ) is the output of the waste household appliance disassembly part network, F is the input feature of G ω (F; ω t ), ω t is the weight of G ω (F; ω t ), and the size of the weight matrix is 2048×10;

[0042] (3) Train the waste household appliance disassembly part recognition model based on sparse representation precision quantization

[0043] Train the waste household appliance disassembly part recognition model based on sparse representation precision quantization. Input the training set X s , and after passing through F θ (X s ; θ t ) and G ω (F; ω t ), complete the training of the recognition model. The expression of the total training loss function L ω (F; ω t ) is: D For:

[0044] L D = L ce + L qe + L bc (11)

[0045] Among them, L qe is the quantization error loss function, L bc is the distribution loss function, L ce is the cross-entropy loss function, and the formula is:

[0046]

[0047] Among them, is the true label of the i-th image data of the j-th category of images, j = 1, 2,..., 12, i = 1, 2,..., N, and N is the number of images of the j-th category, is the predicted probability that image i belongs to category j, Y is the true label of the input image category, P is the predicted probability of the input image category, L qe Calculate the losses of the model parameters before and after quantization. The formula is:

[0048]

[0049] Among them, t = 1, 2,..., T, and T is the number of training times of the waste household appliance disassembly part recognition model, is the quantized weight, and the corresponding formula is:

[0050]

[0051] Among them, round(·) represents taking an integer, and Δ w is the quantization step size, and the formula for Δ w is:

[0052]

[0053] Among them, max(·) represents taking the maximum value, min(·) represents taking the minimum value, and the distribution loss L bc is used to ensure that the output distribution of the quantized model is consistent with that of the unquantized model, and the formula is:

[0054]

[0055] Among them, E is the calculation of the expected value, ‖·‖ is the modulus of the vector, and O S is the output of the layer corresponding to the minimum bit width, and O H is the output of the layer corresponding to the maximum bit width, is the weight sparse representation, and the corresponding formula is:

[0056]

[0057] Among them, ⊙ is the channel multiplication, and H(·) is the unit step function, and the formula is:

[0058]

[0059] Use the gradient descent algorithm to optimize the parameters θ t and ω t in the network for identifying waste household appliance disassembly parts with sparse representation accuracy quantization. The parameter update formula is:

[0060]

[0061] Among them, ω t+1 is the parameter matrix of G ω (F; ω t ) during the (t + 1)-th iteration training, and θ t+1 is the parameter matrix of F θ (X t ; θ t ) during the (t + 1)-th iteration training. represents taking the partial derivative. When t is 300 or more, terminate the training of the recognition model, save the weight parameters of the last training, and complete the training of the recognition model;

[0062] (4) Identify the types of waste household appliance disassembly parts

[0063] After completing the model training, load the weights of the waste household appliance disassembly part recognition model based on sparse representation accuracy quantization, and for the test set Identify the images of 12 types of waste household appliance disassembly parts in to obtain the predicted class labels of waste household appliance disassembly parts and complete the task of identifying waste household appliance disassembly parts.

Claims

1. A method for identifying disassembled parts of waste household appliances based on sparse representation accuracy quantification, characterized in that: The following steps are involved: (1) Collecting data on dismantled parts of used household appliances Obtain a dataset of disassembled parts of used household appliances, including 12 types of disassembled parts images of used household appliances: circuit boards, wires, aluminum, compressors, condensers, transformers, copper pipes, external motors, drain pipes, flame retardants, waste and internal fans, with a total of s images in JPG format. In order to meet the input requirements of the model, the pixels of each image in the dataset of disassembled parts of used household appliances are normalized to [0,1]. The normalized image dataset of disassembled parts of used household appliances is divided into a training set and a test set. The training set is X s , the test set is With X s The number of categories of images of disassembled parts of used household appliances is 12; (2) Constructing a recognition model for disassembled parts of used household appliances Construct a network for disassembling components of used household appliances based on ConvNeXt-Tiny: The network consists of five parts. The first part conv1 consists of a convolution layer with a convolution kernel of 1×1, a batch normalization processing layer, and a GeLU activation function layer, with 96 channels. The second part conv2 consists of two inverted bottleneck modules, each of which consists of a depth-separable convolution layer with a convolution kernel of 1×1, a GeLU activation function layer, and a convolution layer with a convolution kernel of 1×1, with 96 channels. The third part conv3 consists of two inverted bottleneck modules, each of which consists of a convolution kernel of 1× 1 depth separable convolution layer, a GeLU activation function layer, a convolution layer with a convolution kernel of 1×1, and the number of channels is 192. The fourth part conv4 consists of two inverted bottleneck modules, each of which consists of a depth separable convolution layer with a convolution kernel of 1×1, a GeLU activation function layer, and a convolution layer with a convolution kernel of 1×1. The number of channels is 384. The fifth part conv5 consists of two inverted bottleneck modules, each of which consists of a depth separable convolution layer with a convolution kernel of 1×1, a GeLU activation function layer, and a convolution layer with a convolution kernel of 1×1, and the number of channels is 768; X s Define F as the input of the ConvNeXt-T network for disassembling components of used household appliances. θ (X s θ t ) is the feature extractor of the network for dismantling components of used household appliances, t is the number of iterative training, θ t F θ (X s θ t ) has learnable parameters, the parameter matrix size is 512×28×28, and the predictor G ω (F;ω t ) is the output of the network for dismantling components of used household appliances, F is G ω (F;ω t ), ω t G ω (F;ω t ), the weight matrix size is 2048×10; (3) Training a model for identifying disassembled parts of used household appliances based on sparse representation accuracy quantification Training of the recognition model of disassembled parts of used household appliances based on sparse representation accuracy quantization, input training set X s , through F θ (X s θ t ) and G ω (F;ω t ) completes the recognition model training, G ω (F;ω t )’s total training loss function expression L D for: L D =L ce +L qe +L bc (1) Among them, L qe is the quantization error loss function, L bc is the distribution loss function, L ce is the cross entropy loss function, and the formula is: in, is the true label of the i-th image data of the j-th category image, j=1,2,…,12, i=1,2,…,N, N is the number of images of the j-th category, is the predicted probability that image i belongs to category j, Y is the true label of the input image category, P is the predicted probability of the input image category, L qe Calculate the unquantized and quantized model parameter loss using the formula: Among them, t = 1, 2, ..., T, T is the number of training times of the dismantled parts recognition model of used household appliances, is the quantized weight, and the corresponding formula is: Among them, round(·) means taking an integer, Δ w is the quantization step size, Δ w The formula is: Among them, max(·) means taking the maximum value, min(·) means taking the minimum value, and the distribution loss L bc It is used to ensure that the output distribution of the quantized model is consistent with the unquantized model. The formula is: Among them, E is the expected value calculation, ||·|| is the modulus of the vector, O S is the output of the layer corresponding to the minimum bit width, O H is the output of the layer corresponding to the maximum bit width, is the weight sparse representation, and the corresponding formula is: Where ⊙ is the channel multiplication, H(·) is the unit step function, and the formula is: Using gradient descent algorithm to optimize the parameters θ in the sparse representation accuracy quantization network of dismantled components of waste household appliances t and ω t , the parameter update formula is: Among them, ω t+1 G is the t+1th iteration training ω (F;ω t ) parameter matrix, θ t+1 F is the training time of the t+1th iteration θ (X t θ t ), Indicates partial derivative. When t is greater than 300, the recognition model training is terminated, the weight parameters of the last training are saved, and the recognition model training is completed. (4) Identify the types of parts from dismantled waste appliances After completing the model training, load the weights of the model for identifying disassembled parts of used household appliances based on sparse representation accuracy quantization, and The 12 types of disassembled parts images of used household appliances are identified to obtain the predicted category labels of disassembled parts of used household appliances Complete the task of identifying parts from dismantled used household appliances.

Images