Granted Patent B2

US 12,443,828 · App. 18/932,549 · Granted Oct 14, 2025

Inventors: Hyungjun Lee (Daejeon, KR); Hancheol Park (Daejeon, KR)

Assignee: NOTA, INC.

G06N3/0495G06N3/082G06T7/62G06V10/764

Abstract

Claims (60)

1. A method of compressing, by an electronic device, a neural network model for object recognition, the method comprising:

receiving an original model for object recognition trained based on a first data set;

receiving a second data set of an analysis target;

calculating object size information on sizes of objects included in an image of the second data set;

calculating performance decline rates for each object size of the original model according to pruning, based on the calculated object size information;

determining weight information on pruning ratios for each object size in consideration of the performance decline rates for each object size;

determining an optimized pruning ratio for the second data set of at least one layer included in the original model based on the determined weight information; and

generating a compressed neural network model from the original model by performing the pruning on the at least one layer based on the determined optimized pruning ratio, and

wherein the object size information includes information on a first ratio of first objects corresponding to a first size range included in the image and a second ratio of second objects corresponding to a second size range included in the image,

the first ratio is calculated based on number of the first objects in the image and total number of the total objects in the image including the first objects and the second objects, and

the second ratio is calculated based on number of the second objects in the image and the total number.

2. The method of claim 1 , wherein the calculating of the object size information further includes:

receiving the number of pixels corresponding to the objects included in the image of the second data set;

when the number of pixels corresponding to the object is larger than or equal to the predetermined number of reference pixels, classifying the object as the first object corresponding to the first size range, and when the number of pixels corresponding to the object is smaller than the number of reference pixels, classifying the object as the second object corresponding to the second size range; and

calculating the object size information based on the number of objects classified as the first object and the number of objects classified as the second object.

3. The method of claim 1 , wherein the determining of the performance decline rates for each object size further includes:

receiving first graph information indicating sensitivity of the original model to the pruning ratio corresponding to the first size range and second graph information indicating the sensitivity of the original model relative to the pruning ratio corresponding to the second size range;

calculating a first target performance value for the first size range based on the first ratio, and calculating a second target performance value for the second size range based on the second ratio; and

calculating the performance decline rates for each object size based on the first target performance value and the first graph information, and the second target performance value and the second graph information.

4. The method of claim 3 , wherein the first graph information and the second graph information are generated during the pruning after training of the original model using the first data set is completed.

5. The method of claim 3 , wherein, when the first ratio is smaller than the second ratio, the first target performance value is calculated as a higher value than the second target performance value.

6. The method of claim 3 , wherein the determining of the performance decline rates for each object size further includes:

calculating a first maximum performance value from the first graph information and a second maximum performance value from the second graph information; and

calculating first performance decline rates corresponding to objects in the first size range based on a difference between the first maximum performance value and the first target performance value, and calculating second performance decline rates corresponding to objects in the second size range based on a difference between the second maximum performance value and the second target performance value, and

the determining of the weight information includes calculating first weights corresponding to the objects in the first size range included in the weight information based on the first performance decline rate and calculating second weights corresponding to the objects in the second size range included in the weight information based on the second performance decline rate.

7. The method of claim 6 , wherein the determining of the optimized pruning ratio further includes:

calculating a first pruning ratio corresponding to the first target performance value from the first graph information and calculating a second pruning ratio corresponding to the second target performance value from the second graph information;

calculating an adjusted first pruning ratio by assigning the first weight to the first pruning ratio, and calculating an adjusted second pruning ratio by assigning the second weight to the second pruning ratio; and

calculating the optimized pruning ratio based on the adjusted first pruning ratio or the adjusted second pruning ratio.

8. The method of claim 1 , wherein the original model includes a first layer and a second layer located between an input layer and an output layer of the original model, and the first layer is located relatively closer to the input layer than the second layer,

the determining of the optimized pruning ratio further includes calculating a first optimized pruning ratio for the first layer and calculating a second optimized pruning ratio for the second layer, and

the generating of the compressed neural network model from the original model includes performing the pruning on the first layer based on the first optimized pruning ratio and performing the pruning on the second layer based on the second optimized pruning ratio to generate the compressed neural network model.

9. The method of claim 1 , wherein the pruning ratios for each object size are determined for each object size based on graph information indicating the sensitivity of the original model, and a target performance value according to the pruning, and

include a first pruning ratio for first objects corresponding to a first size range included in the image and a second pruning ratio for second objects corresponding to a second size range included in the image.

10. An electronic device, comprising:

a transceiver configured to receive an original model for object recognition on which training is completed; and

a processor configured to perform compressing on the original model,

wherein the processor is configured to receive an original model for object recognition trained based on a first data set, receive a second data set of an analysis target, calculate object size information on sizes of objects included in an image of the second data set, calculate performance decline rates for each object size of the original model according to pruning, based on the calculated object size information, determine weight information on pruning ratios for each object size in consideration of the performance decline rates for each object size, determine an optimized pruning ratio for the second data set of at least one layer included in the original model based on the determined weight information, and generate a compressed neural network model from the original model by performing the pruning on the at least one layer based on the determined optimized pruning ratio, and

wherein the object size information includes information on a first ratio of first objects corresponding to a first size range included in the image and a second ratio of second objects corresponding to a second size range included in the image,

the first ratio is calculated based on number of the first objects in the image and total number of the total objects in the image including the first objects and the second objects, and

the second ratio is calculated based on number of the second objects in the image and the total number.

11. A method of recognizing, by an electronic device, an object included in a data set to be analyzed using a neural network model, the method comprising:

receiving the data set; and

inputting the received data set to a compressed neural network model to generate an output value corresponding to a value identifying an object included in an image of the data set,

wherein the compressed neural network model is compressed from an original model using an optimized pruning ratio determined for at least one layer of the original model based on ratios for each object size of objects included in the image of the data set, and

the optimized pruning ratio is calculated for the data set based on the determined weight by calculating performance decline rates for each object size of the original model according to pruning based on ratios for each object size and determining weights for pruning ratios for each object size in consideration of the performance decline rates for each object size, and

wherein the ratios for each object size include a first ratio of first objects corresponding to a first size range included in the data set and a second ratio of second objects corresponding to a second size range included in the data set,

the first ratio is calculated based on number of the first objects in the image and total number of the total objects in the image including the first objects and the second objects, and

the second ratio is calculated based on number of the second objects in the image and the total number.

12. The method of claim 11 , wherein the original model is a model pre-trained using a data set different from the data set, and

the optimized pruning ratio is determined by applying a target performance value calculated for each object size from ratios for each object size of the data set to pruning sensitivity information of the at least one layer calculated for each object size using the other data set.

13. The method of claim 12 , wherein the optimized pruning ratio is determined by applying the target performance value to the pruning sensitivity information and weight-summing the pruning ratios for each object size determined for each object size with a different weight for each object size,

the performance decline rates for each object size are determined based on the pruning sensitivity information and the target performance value, and

the weight is determined to be larger as the performance decline rates for each object size increase.

14. The method of claim 13 , wherein the performance decline rates for each object size are defined based on a difference between a maximum performance value calculated from the pruning sensitivity information and the target performance value, and the pruning ratios for each object size.

15. The method of claim 12 , wherein

when the first ratio is smaller than the second ratio, the target performance value determined for the first layer related to the first ratio is determined to be larger than the target performance value determined for the second layer related to the second ratio.

16. The method of claim 15 , wherein the ratios for each object size are calculated based on the number of objects corresponding to the first size range and the number of objects corresponding to the second size range classified based on the number of pixels corresponding to the objects in the data set and the predetermined number of reference pixels.

17. The method of claim 11 , wherein the pruning ratios for each object size are determined for each object size based on pruning sensitivity information indicating the sensitivity of the original model, and a target performance value according to the pruning, and

include a first pruning ratio for the first objects corresponding to a first size range included in the image and a second pruning ratio for the second objects corresponding to a second size range included in the image.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (13)