US 11,955,272 B1
Method for generating object detector based on deep learning capable of detecting extended object class and detector generation device using the same
Kye Hyeon Kim, Suwon-si (KR)
Assigned to SUPERB AI CO., LTD., Seoul (KR)
Filed by Superb AI Co., Ltd., Seoul (KR)
Filed on Oct. 27, 2023, as Appl. No. 18/384,731.
Claims priority of application No. 10-2022-0149470 (KR), filed on Nov. 10, 2022.
Int. Cl. H01F 27/40 (2006.01); G01N 33/00 (2006.01)
CPC H01F 27/402 (2013.01) [G01N 33/0034 (2013.01); H01F 2027/404 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method for generating an object detector based on deep learning capable of detecting an extended object class, comprising steps of:
(a) on condition that a first object detector is generated, wherein the first object detector has been trained to detect one or more first objects in a first input image by using a first training data set generated by annotation of first object classes of the first objects in each of first source images containing at least one of the first objects, in response to acquiring second source images containing one or more second objects, a detector generation device generating a second training data set by annotation of second object classes of the second objects for each of the second source images, and generating a second object detector trained to detect the second objects in a second input image by using the second training data set;
(b) the detector generation device (i) (i-1) inputting the first source images of the first training data set into the second object detector, thereby instructing the second object detector to detect the second objects contained in each of the first source images and thus to output second object detection results, and (i-2) generating a third training data set containing the first source images in which the first object classes and the second object classes are annotated by annotation of the second object classes for the second objects detected in each of the first source images by referring to the second object detection results, and (ii) (ii-1) inputting the second source images of the second training data set into the first object detector, thereby instructing the first object detector to detect the first objects contained in each of the second source images and thus to output first object detection results, and (ii-2) generating a fourth training data set containing the second source images in which the first object classes and the second object classes are annotated by annotation of the first object classes for the first objects detected in each of the second source images by referring to the first object detection results; and
(c) the detector generation device generating an integrated object detector trained to detect the first objects and the second objects in a third input image by using the third training data set and the fourth training data set.