US 2026/0279000 A1 · Precision-Quantized Sparse Representation Recognition Method for Disassembled Parts of Waste Electrical Appliance

Abstract

A method for recognizing disassembled components of waste electrical appliances based on sparse representation with precision quantization is disclosed. The method addresses instability in decision-related features and weight allocation during the quantization process of deep learning models. A ConvNeXt-Tiny network is used to extract deep features from images of disassembled components. A quantization error loss function is applied to optimize the training of weights, while a distribution loss function freezes unstable bit-width parameters during training. This approach enables accurate classification of disassembled components and improves overall recognition accuracy. In practical disassembly scenarios, the method effectively mitigates quantization-related instability and achieves efficient classification and identification of component types, providing technical support for the recycling of waste electrical appliances.

A precision-quantized sparse representation recognition method for disassembled parts of waste electrical appliance, is implemented via the following device structure:

an integrated edge computing recognition terminal device includes an industrial grade high resolution camera, a gigabit ethernet port, a visual pre-processing module, a neural processing unit, a capacitive touchscreen, and a data server:
the industrial grade high resolution camera is used to capture images of disassembled parts on the waste electrical appliance disassembly line;
the gigabit ethernet port is used to transmit the captured images to the subsequent computation unit in real time;
the visual pre-processing module is used to adjust the brightness, contrast, and sharpness of the captured images;
the neural network processor is used to perform image recognition;
the capacitive touch screen is used to display the recognition results of the disassembled parts;
the data server is used to store the recognition results;
the workflow of the integrated edge computing recognition terminal device is as follows:
first, the industrial grade high resolution camera captures and transmits images of the disassembled parts at the edge computing terminal via an ethernet data interface;
second, the visual pre-processing module enhances the quality of the captured images, and the neural processing unit outputs the image recognition results;
third, the category labels of the disassembled parts are displayed on the capacitive touch screen, and the category labels of the disassembled parts are transformed into a structured form and stored in the data server, with the specific recognition method comprising:
(1) get a dataset containing images of disassembled parts of waste electrical appliance categorized into 12 types:
including circuit boards, wires, aluminum, compressors, condensers, transformers, copper tubes, external motors, drains, flame-retardant, waste clutter, and internal fans;
normalize the pixel values of each image in the disassembled parts of waste electrical appliance dataset to the range [0,1], to meet recognition model requirements;
the normalized image dataset of disassembled parts of waste electrical appliance is divided into a training set and a test set, with the training set being Xs and the test set being {circumflex over (X)}s;
s is the total number of images with disassembled parts of waste electrical appliance;
the total number of {circumflex over (X)}s images is 30% of s;
the disassembled parts of waste electrical appliances in both {circumflex over (X)}s and Xs consist of 12 categories;
(2) construct the disassembled parts recognition model of waste electrical appliances with ConvNeXt-Tiny:
the model consists of five parts, the first part conv1 consists of a 1×1 convolutional layer, a batch normalization layer, and a GeLU activation function layer, with a channel count of 96;
the second part conv2 includes two inverted bottleneck modules, each consisting of a 1×1 depthwise separable convolutional layer, a GeLU activation function layer, and a 1×1 convolutional layer, with a channel count of 96;
the third part conv3 contains two inverted bottleneck modules, each consisting of a 1×1 depthwise separable convolutional layer, a GeLU activation function layer, and a 1×1 convolutional layer, with a channel count of 192;
the fourth part conv4 is composed of two inverted bottleneck modules, each consisting of a 1×1 depthwise separable convolutional layer, a GeLU activation function layer, and a 1×1 convolutional layer, with a channel count of 384;
the fifth part conv5 incorporates two inverted bottleneck modules, each consisting of a 1×1 depthwise separable convolutional layer, a GeLU activation function layer, and a 1×1 convolutional layer, with a channel count of 768;
Xs is the input for the disassembled parts recognition model of waste electrical appliance with ConvNeXt-T;
define Fθ (Xs;
θt) as the feature extractor of the disassembled parts recognition model of waste electrical appliance;
t is the number of training epochs;
θt is the learnable parameters of Fθ (Xs;
θt) with a parameter matrix size of 512×28×28;
Gω (F;
ωt) is the output of the disassembled parts recognition model of waste electrical appliance;
F is the input features of Gω (F;
ωt), ωt is the weight of Gω (F;
ωt) with a weight matrix size of 2048×10;
(3) the recognition model of disassembled parts of waste electrical appliances is trained with precision-quantized sparse representation;
the recognition model is trained with input training set Xs;
the recognition model training is completed after processing through Fθ (Xs;
θt) and Gω (F;
ωt);
the training total loss function LD for Gω (F;
ωt) is:
L
D
=
L
ce
+
L
qe
+
L
bc
,
where Lqe is the quantization error loss function;
Lbc is the distribution loss function;
the cross-entropy loss function Lce is calculated as:
L
ce
=
-
∑
j
=
1
12
∑
i
=
1
N
Y
j
(
i
)
⁢
log
⁢
P
j
(
i
)
,
where Yj(i) is the i-th image data true label of the j-th image category, j=1, 2, . . . , 12, i=1, 2, . . . , N, N is the number of the j-th image category;
P
j
(
i
)
is the predicted probability of image i belonging to category j;
Y is the input image category true label;
P is the input image category predicted probability;
the unquantized and quantized model parameter losses Lqe is calculated as:
L
qe
=
1
N
⁢
∑
t
=
1
N
(
ω
t
-
ω
^
t
)
2
,
where t=1, 2, . . . , T, T is the total number of training epochs for the disassembled parts recognition model of waste electrical appliance;
the weight after quantization ωt is calculated as:
ω
^
t
=
round
(
ω
t
Δ
ω
)
·
Δ
w
,
where round(·) is the integer rounding operation;
the quantization step Δw is calculated as:
Δ
ω
=
max
⁡
(
ω
t
)
-
min
⁡
(
ω
t
)
2
8
-
1
,
where max(·) denotes taking the maximum value and min(·) denotes taking the minimum value;
distribution loss Lbc is used to ensure that the output distribution of the quantised model is consistent with the unquantised model;
Lbc is calculated as:
L
bc
=
E
[
▯
⁢
F
θ
(
X
t
;
θ
t
)
⁢
(
O
S
-
O
H
)
⁢
▯
+
0.01
▯
⁢
W
^
instability
t
⁢
▯
]
,
where E[·] denotes the expectation function;
is the mode of the vector;
OS is the output of the layer corresponding to the minimum bit width;
OH is the output of the layer corresponding to the maximum bit width;
the weighted sparse representation Ŵinstabilityt is calculated as:
W
^
instability
t
=
max
⁢
∑
t
=
1
300
1
2
8
-
1
⁢

ω
t

2
2
⁢
▯
⁢
H
⁡
(
ω
t
)
,
where □ is the channel multiplication and the unit step function H(·) is:
H
⁡
(
ω
t
)
=
{
0
ω
t
<
0
1
ω
t
≥
0
,
the parameters θt and ωt in the disassembled parts recognition model of waste electrical appliances are optimized using the gradient descent algorithm;
the parameter updating formulae are:
θ
t
+
1
=
θ
t
-
0.01
∂
L
D
∂
θ
t
,
ω
t
+
1
=
ω
t
-
0.01
∂
L
D
∂
ω
t
,
where ωt+1 is the parameter matrix of Gω (F;
ωt) at the (t+1)-th iteration of training;
θt+1 is the parameter matrix of Fθ (Xt;
θt) at the (t+1)-th iteration of training;
& denotes the partial derivatives;
at t=300, terminate the training of the recognition model and save the weight parameters of the last training;
(4) upon completion of model training, load the weights for the disassembled parts recognition model of waste electrical appliance with precision-quantized sparse representation recognition;
use the trained model to recognize 12 categories of disassembled parts in the test set {circumflex over (X)}s, and obtain the prediction labels Ŷs.

295 more elements

Precision-Quantized Sparse Representation Recognition Method for Disassembled Parts of Waste Electrical Appliance

Life of this patent

Precision-Quantized Sparse Representation Recognition Method for Disassembled Parts of Waste Electrical Appliance

Life of this patent

Abstract

Claims1