METHOD AND DEVICE FOR DEFECT DETECTION
A method and a device for defect detection are provided. The method includes: obtaining a defect eigenvector of an image to be detected; calculating a similarity score of the image to be detected for each known defect type according to the defect eigenvector; and performing defect classification on the image to be detected according to the similarity score. The similarity score between the defect eigenvector of the image to be detected and each known defect type is calculated, and then it is possible to distinguish known defects from unknown defects and accurately classify the known defects according to the similarity scores, so as to effectively detect the defects.
1 . A method for defect detection, comprising:
obtaining a defect eigenvector of an image to be detected;
calculating a similarity score of the image to be detected for each known defect type according to the defect eigenvector; and
performing defect classification on the image to be detected according to the similarity score.
2 . The method according to claim 1 , wherein the calculating a similarity score of the image to be detected for each known defect type according to the defect eigenvector comprises:
mapping the defect eigenvector into a trained eigenvector space, wherein the trained eigenvector space comprises a distribution position of a defect eigenvector of the known defect type; and
calculating a distance between the defect eigenvector and the defect eigenvector of the known defect type, so as to obtain the similarity score of the image to be detected for each known defect type.
3 . The method according to claim 2 , wherein the performing defect classification on the image to be detected according to the similarity scores comprises:
when a maximum similarity score in similarity scores is not less than a similarity threshold, outputting a defect type corresponding to the maximum similarity score; or
when the similarity scores are all less than the similarity threshold, outputting an unknown defect type.
4 . The method according to claim 3 , further comprising:
outputting the similarity score corresponding to the defect type.
5 . The method according claim 1 , wherein the obtaining a defect eigenvector of an image to be detected comprises:
obtaining the defect eigenvector by an encoding module.
6 . The method according to claim 5 , wherein the encoding module comprises a plurality of encoding sublayers, the plurality of encoding sublayers being connected by a maximum pooling layer; and
wherein the obtaining the defect eigenvector by an encoding module comprises:
inputting the image to be detected to the encoding module, so as to obtain feature maps of sublayers that correspond to the plurality of encoding sublayers; and
converting a feature map of a final sublayer in the feature maps of the sublayers into the defect eigenvector.
7 . The method according to claim 5 , further comprising:
obtaining defect position information of the image to be detected by a decoding module.
8 . The method according to claim 7 , wherein the decoding module comprises a plurality of decoding sublayers, the plurality of decoding sublayers being connected through deconvolution and corresponding to the plurality of encoding sublayers; and
wherein the obtaining defect position information of the image to be detected by a decoding module comprises:
inputting the defect eigenvector to the decoding module, so as to obtain feature maps corresponding to the plurality of decoding sublayers,
wherein a feature map of a final sublayer in the plurality of decoding sublayers comprises the defect position information.
9 . The method according to claim 6 , wherein before the calculating a similarity score of the image to be detected for each known defect type according to the defect eigenvector, the method further comprises:
obtaining a plurality of known defect sample images; and
training an eigenvector space by using the plurality of known defect sample images, so as to obtain the trained eigenvector space.
10 . The method according to claim 9 , wherein the training an eigenvector space by using the plurality of known defect sample images, so as to obtain the trained eigenvector space comprises:
obtaining known defect eigenvectors of the plurality of known defect sample images by the encoding module; and
minimizing a distance between the known defect eigenvectors of a same type and maximizing a distance between the known defect eigenvectors of different types in a metric learning function.
11 . A device for defect detection, comprising:
an obtaining unit, configured to obtain a defect eigenvector of an image to be detected; and
a processing unit, configured to calculate a similarity score of the image to be detected for each known defect type according to the defect eigenvector, and to perform defect classification on the image to be detected according to the similarity score.
12 . The device according to claim 11 , wherein the processing unit is further configured to:
map the defect eigenvector into a trained eigenvector space, wherein the trained eigenvector space comprises a distribution position of a defect eigenvector of the known defect type; and
calculate a distance between the defect eigenvector and the defect eigenvector of the known defect type, so as to obtain the similarity score of the image to be detected for each known defect type.
13 . The device according to claim 12 , further comprising an output unit, wherein the output unit is configured to:
when a maximum similarity score in similarity scores is not less than a similarity threshold, output a defect type corresponding to the maximum similarity score; or
when the similarity scores are all less than the similarity threshold, output an unknown defect type.
14 . The device according to claim 13 , wherein the output unit is further configured to output the similarity score corresponding to the defect type.
15 . The device according to claim 11 , wherein the obtaining unit is further configured to obtain the defect eigenvector by an encoding module.
16 . The device according to claim 15 , wherein the encoding module comprises a plurality of encoding sublayers, the plurality of encoding sublayers being connected by a maximum pooling layer; and wherein the processing unit is configured to:
input the image to be detected to the encoding module, so as to obtain feature maps of sublayers that correspond to the plurality of encoding sublayers; and
convert a feature map of a final sublayer in the feature maps of the sublayers into the defect eigenvector.
17 . The device according to claim 15 , wherein the obtaining unit is further configured to obtain defect position information of the image to be detected by a decoding module.
18 . The device according to claim 17 , wherein the decoding module comprises a plurality of decoding sublayers, the plurality of decoding sublayers being connected through deconvolution and corresponding to the plurality of encoding sublayers; and wherein the processing unit is further configured to:
input the defect eigenvector to the decoding module, so as to obtain feature maps corresponding to the plurality of decoding sublayers,
wherein a feature map of a final sublayer in the plurality of decoding sublayers comprises the defect position information.
19 . An apparatus for defect detection, comprising a processor and a memory, wherein the memory is configured to store a program, and the processor is configured to call the program from the memory and run the program to perform the method for defect detection according to claim 1 .
20 . A non-transitory computer-readable storage medium, storing a computer program, wherein the computer program, when run on a computer, causes the computer to perform the method for defect detection according to claim 1 .