Machine learning method and apparatus
A machine learning method for learning how to form bounding boxes, performed by a machine learning apparatus, includes extracting learning images including a target object among a plurality of learning images included in a learning database, generating additional learning images in which the target object is rotated from the learning images including the target object, and updating the learning database using the additional learning images.
1. A machine learning method for learning how to form bounding boxes, performed by a machine learning apparatus, the machine learning method comprising:
extracting learning images including a target object among a plurality of learning images included in a learning database;
obtaining information on a distribution of aspect ratios of bounding boxes of the target object in the learning images including the target object;
generating additional learning images in which the target object is rotated from the learning images including the target object; and
updating the learning database using the additional learning images.
2. The machine learning method according to claim 1 , wherein, in the generating additional learning images, a rotation angle of the target object included in each of the additional learning images is determined based on the distribution of aspect ratios of bounding boxes of the target object.
3. The machine learning method according to claim 2 , wherein, in the generating additional learning images, the additional learning images are generated so that a distribution of aspect ratios of bounding boxes of the target object included in the additional learning images follows the distribution of aspect ratios of bounding boxes of the target object in the learning images including the target object.
4. The machine learning method according to claim 1 , wherein, in the updating, a bounding box of the target object included in each of the additional learning images is reformed, and information on the reformed bounding box are added to the learning database as being labeled to each of the additional learning images.
5. The machine learning method according to claim 1 , further comprising generating a bounding box formation model using the updated learning database.
6. A machine learning method for learning how to form bounding boxes, performed by a machine learning apparatus, the machine learning method comprising:
extracting learning images including a target object among a plurality of learning images included in a learning database;
determining an image type of the target object;
determining learning parameters for the learning images including the target object based on the image type of the target object; and
generating a bounding box formation model based on the learning database and the learning parameters,
wherein, in the determining an image type of the target object, information on specifications of bounding boxes of the target object in the learning images including the target object is obtained, and the image type of the target object is determined based on the information on specifications of bounding boxes.
7. The machine learning method according to claim 6 , wherein, in the determining an image type of the target object, the image type of the target object is determined as one of a horizontal type, a vertical type, and a normal type based on aspect ratios of the bounding boxes of the target object.
8. The machine learning method according to claim 7 ,
wherein the image type of the target object is determined as the horizontal type when both of Equations 1 and 2 are satisfied, determined as the vertical type when both of Equations 1 and 2 are not satisfied, or determined as the normal type when only one of Equations 1 and 2 is satisfied, and
wherein Equation 1 is ‘Aspect_ratio_mean>Th1’, Equation 2 is ‘Aspect_ratio_max/Aspect_ratio_min>Th2’, Aspect_ratio_mean denotes an average value of aspect ratios of the bounding boxes of the target object, Aspect_ratio_max denotes a maximum value among the aspect ratios of the bounding boxes of the target object, Aspect_ratio_min denotes a minimum value among the aspect ratios of the bounding boxes of the target object, Th1 denotes a first reference value, and Th2 denotes a second reference value.
9. A machine learning apparatus for learning how to form bounding boxes, comprising a processor and a memory storing at least one instruction executed by the processor, wherein the at least one instruction is configured to:
extract learning images including a target object among a plurality of learning images included in a learning database;
obtain information on a distribution of aspect ratios of bounding boxes of the target object in the learning images including the target object;
generate additional learning images in which the target object is rotated from the learning images including the target object; and
update the learning database using the additional learning images.
10. The machine learning apparatus according to claim 9 , wherein the at least one instruction is further configured to determine a rotation angle of the target object included in each of the additional learning images based on the distribution of aspect ratios of bounding boxes of the target object.
11. The machine learning apparatus according to claim 10 , wherein the at least one instruction is further configured to generate the additional learning images so that a distribution of aspect ratios of bounding boxes of the target object included in the additional learning images follows the distribution of aspect ratios of bounding boxes of the target object in the learning images including the target object.
12. The machine learning apparatus according to claim 9 , wherein the at least one instruction is further configured to reform a bounding box of the target object included in each of the additional learning images, and add information on the reformed bounding box to the learning database as being labeled to each of the additional learning images.
13. The machine learning apparatus according to claim 9 , wherein the at least one instruction is further configured to generate a bounding box formation model using the updated learning database.
14. The machine learning apparatus according to claim 9 , wherein the at least one instruction is further configured to determine an image type of the target object, determine learning parameters for the learning images including the target object based on the image type of the target object, and generate a bounding box formation model based on the learning database and the learning parameters.
15. The machine learning apparatus according to claim 14 , wherein the at least one instruction is further configured to obtain information on specifications of bounding boxes of the target object in the learning images including the target object, and determine the image type of the target object based on the information on specifications of bounding boxes.
16. The machine learning apparatus according to claim 15 , wherein the at least one instruction is further configured to determine the image type of the target object as one of a horizontal type, a vertical type, and a normal type based on aspect ratios of the bounding boxes of the target object.
17. The machine learning apparatus according to claim 16 ,
wherein the at least one instruction is further configured to determine the image type of the target object as the horizontal type when both of Equations 1 and 2 are satisfied, determine the image type of the target object as the vertical type when both of Equations 1 and 2 are not satisfied, or determine the image type of the target object as the normal type when only one of Equations 1 and 2 is satisfied, and
wherein Equation 1 is ‘Aspect_ratio_mean>Th1’, Equation 2 is ‘Aspect_ratio_max/Aspect_ratio_min>Th2’, Aspect_ratio_mean denotes an average value of aspect ratios of the bounding boxes of the target object, Aspect_ratio_max denotes a maximum value among the aspect ratios of the bounding boxes of the target object, Aspect_ratio_min denotes a minimum value among the aspect ratios of the bounding boxes of the target object, Th1 denotes a first reference value, and Th2 denotes a second reference value.