Subtle defect detection method based on coarse-to-fine strategy

1. A subtle defect detection method based on coarse-to-fine strategy, comprising:

(S1) acquiring data of an image to be detected via a charge-coupled device (CCD) camera;

(S2) constructing a defect area location network and preprocessing the image to be detected to initially determine a defect position;

(S3) constructing a defect point detection network; and training the defect point detection network by using a defect segmentation loss function; and

(S4) according to the defect position initially determined in step (S2), subjecting a subtle defect in the image to be detected to quantitative extraction and segmentation by using the defect point detection network;

wherein the defect point detection network comprises a backbone network comprising six stages, a bidirectional feature pyramid network, a classification network and a regression network;

an input image of the backbone network is an image output by the defect area location network, and the backbone network is configured to extract a defect feature of the input image;

in the six stages, a first stage comprises a convolutional layer and a 7×7 convolution kernel;

a second stage comprises a 3×3 max-pooling layer and a first dense block;

the second stage further comprises alternating 1×1 and 3×3 convolution kernels;

a third stage is composed of a second dense block; a fourth stage is composed of a third dense block structurally different from the second dense block;

the third stage and the fourth stage are configured to accelerate transmission of the defect feature and improve utilization of a defect feature image;

a fifth stage is composed of two dilated bottleneck layers to capture subtle target defect features; and

a sixth stage is composed of a dilated bottleneck layer to avoid loss of the subtle target defect features.

2. The subtle defect detection method of claim 1 , wherein step (S2) comprises:

(21) constructing the defect area location network, wherein the defect area location network comprises a backbone network, a classification network and a regression network;

(22) inputting the image to be detected into the backbone network; and extracting defect information from the image to be detected via a 1×1 convolution kernel and a 3×3 convolution kernel;

(23) inputting the defect information to the classification network and the regression network to initially determine the defect position; wherein the classification network is configured to predict whether anchor boxes have the same defect feature; and the regression network is configured to predict a position of each of the anchor boxes.

3. The subtle defect detection method of claim 2 , wherein the classification network and the regression network share a feature weight at the same level; wherein first five layers of a backbone in the backbone network are composed of four convolutional layers and one pooling layer.

4. The subtle defect detection method of claim 1 , wherein the bidirectional feature pyramid network is configured to perform fusion feature mapping on an input defect feature through steps of:

acquiring information of different defect features through bidirectional connection; and balancing the defect features at different layers by variable-weighted feature fusion through the following equation:

O

=

∑

i

1

w

i

ε

+

∑

j

⁢

w

j

·

I

i

;

wherein O is an output feature of the bidirectional feature pyramid network; w i is a guaranteed variable weight, and w i ≥0; ε is a parameter that avoids a denominator from being zero; and l i represents a feature from an i-th layer.

5. The subtle defect detection method of claim 1 , wherein the classification network is configured to predict the defect position; and the regression network is configured to perform defect location and regression, and output a defect identification-location-detection image.

6. The subtle defect detection method of claim 5 , wherein the classification network and the regression network each comprises two convolution kernels; and the classification network and the regression network share a common input feature mapping as fusion feature mapping.

7. The subtle defect detection method of claim 1 , wherein in step (S3), the defect segmentation loss function is used to train a precision of the defect point detection network;

wherein the defect segmentation loss function comprises semantic segmentation loss L ss , edge loss L e , a first regularization loss function and a second regularization loss function;

the semantic segmentation loss L ss is configured to predict a semantic segmentation f by using standard cross entropy (CE) loss, and the edge loss L e is configured to predict a feature mapping s by using standard binary cross entropy (BCE) loss; wherein the semantic segmentation loss L ss is defined as follows:

L ss =λ 1 L CE ({circumflex over (ƒ)},ƒ);

the edge loss L e is defined as follows:

L e =λ 2 L BCE ( s,ŝ );

wherein {circumflex over (ƒ)} and ŝ are defect labels; λ 1 and λ 2 are two balance parameters, and λ 1 and λ 2 ∈[0.1].

8. The subtle defect detection method of claim 7 , wherein last two parts of the defect segmentation loss function are the first regularization loss function and the second regularization loss function;

the first regularization loss function is configured to avoid a mismatch between a defect edge and a predicted edge, defined as follows:

L

r

⁢

1

=

λ

3

⁢

∑

p

+

❘

"\[LeftBracketingBar]"

ζ

⁡

(

p

+

)

-

ζ

ˆ

(

p

+

)

❘

"\[RightBracketingBar]"

;

wherein ζ is a confidence value indicating whether a pixel belongs to a value of the defect edge; p + is set of predicted pixel coordinates; and {circumflex over (ζ)} is a similarity value;

the second regularization loss function is configured to match semantic prediction by using edge prediction to prevent overfitting, defined as follows:

L

r

⁢

2

=

λ

4

⁢

∑

k

,

p

1

s

,

p

[

log

⁢

p

⁡

(

y

p

k

❘

r

,

s

)

]

;

 and

1 s,p ={1: s>thrs};

wherein p indicates a pixel set; k indicates a label set; 1 s,p is an indicator function; thrs is a threshold; λ 3 and λ 4 are two balance parameters to optimize segmentation performance.