CN-122023858-A - Retraining method and device for detection model and non-transitory readable storage medium

Abstract

A retraining method and device for detecting model and a non-transient computer readable storage medium. The method comprises the steps of obtaining accuracy information of an object detection model, retraining the object detection model, and estimating that the accuracy information is abnormal according to an accuracy supervision model.

Inventors

• HUANG YUANFU

Assignees

• 纬创资通股份有限公司

Dates

Publication Date

20260512

Application Date

20241218

Priority Date

20241111

Claims (20)

1. 1. A method of retraining a test model, performed by an computing device, the method comprising: Obtaining accuracy information of an object detection model, and Retraining the object detection model, and estimating that the accuracy information is abnormal according to an accuracy supervision model.
2. 2. The method of retraining a detection model of claim 1, further comprising: obtaining multiple product photographs obtained by photographing multiple products in the same product type, and Judging whether the product photos belong to one identification type in a plurality of screening types by using the object detection model; The accuracy information of the object detection model is the accuracy information of the identification type of the product photos.
3. 3. The method of claim 2, wherein the screening types include at least one of a missing part type, a bad type, a low tin type, an offset type, a short circuit type, and an error type.
4. 4. The method of retraining a test model according to claim 1, wherein the step of retraining the object test model to estimate that the accuracy information is abnormal based on the accuracy supervision model comprises: estimating whether the accuracy information is normal or abnormal by using the accuracy supervision model, and When the accuracy information is abnormal, a model training program is triggered to retrain the object detection model.
5. 5. The method of retraining a test model as recited in claim 4, wherein estimating whether the accuracy information is normal or abnormal using the accuracy supervision model comprises: Analyzing the accuracy information by using the accuracy supervision model to generate accuracy estimation information; Calculating a gap between the accuracy information and the accuracy estimation information, and When the difference is greater than a threshold, the accuracy information is determined to be abnormal, and when the difference is not greater than the threshold, the accuracy information is determined to be normal.
6. 6. The method of claim 1, wherein the accuracy monitor model is a classification model, an encoder-decoder model or a recurrent neural network.
7. 7. The method of retraining a detection model as recited in claim 1, wherein the accuracy supervision model converges with triplet loss during training.
8. 8. The method of claim 1, wherein the accuracy information is time-ordered data comprising a plurality of time-point accuracy data.
9. 9. The method of claim 8, wherein each of the time-point accuracy data includes a miss rate and a false kill rate.
10. 10. The method of claim 1, wherein the object detection model after retraining has more screening type interpretation capabilities than the object detection model prior to retraining.
11. 11. A non-transitory computer readable storage medium storing a plurality of instructions loaded to perform the method of retraining a detection model according to any one of claims 1 to 10.
12. 12. A retraining apparatus for a detection model, comprising: an input unit for obtaining accuracy information of an object detection model, and And the operation unit retrains the object detection model, and estimates that the accuracy information is abnormal according to an accuracy supervision model.
13. 13. The device for retraining a test model according to claim 12, wherein the accuracy information of the object test model is the accuracy information of the object test model for identifying a plurality of product photos belonging to an identification type among a plurality of screening types, the product photos being obtained by photographing a plurality of products in a same product type, respectively.
14. 14. The device for retraining a model as claimed in claim 12, wherein the computing unit uses the accuracy monitor model to estimate whether the accuracy information is normal or abnormal, and triggers a model training procedure to retrain the object detection model when the accuracy information is abnormal.
15. 15. The device of claim 14, wherein the computing unit analyzes the accuracy information using the accuracy supervision model to generate an accuracy estimate, and calculates a gap between the accuracy information and the accuracy estimate, wherein the accuracy information is determined to be abnormal when the gap is greater than a threshold, and the accuracy information is determined to be normal when the gap is not greater than the threshold.
16. 16. The test model retraining apparatus of claim 12, wherein the accuracy monitor model is a classification model, an encoder-decoder model, or a recurrent neural network.
17. 17. The test model retraining apparatus as recited in claim 12, wherein the accuracy supervision model converges with triplet loss during training.
18. 18. The test model retraining apparatus as recited in claim 12, wherein the accuracy information is time-ordered data including a plurality of time-point accuracy data.
19. 19. The test model retraining apparatus of claim 18, wherein each of the time-point accuracy data includes a miss rate and a false kill rate.
20. 20. The test model retraining apparatus of claim 12, wherein the object test model after retraining has more screening type interpretation capabilities than the object test model prior to retraining.

Description

Retraining method and device for detection model and non-transitory readable storage medium Technical Field The present invention relates to a model training method, and more particularly, to a retraining method and apparatus for detecting a model of an object, and a non-transitory computer readable storage medium. Background Currently, industrial manufacturing often uses visual inspection techniques to identify whether a product is defective. However, as time progresses, more and more test data are generated, so that the test model cannot identify some data which are never seen, and the identification accuracy exceeds the standard range of the production line, and the production quality is reduced. As this occurs, the detection model will need to be retrained to correspond to the new data environment. However, on the production line, it is difficult to continuously observe whether the detection model is effectively degraded at any time, and thus it is difficult to efficiently find a point of time at which retraining is required. Disclosure of Invention An embodiment of the present invention provides a retraining method for a detection model, which is executed by an operation device, and includes obtaining accuracy information of an object detection model, retraining the object detection model, and estimating that the accuracy information is abnormal according to an accuracy supervision model. An embodiment of the present invention provides a non-transitory computer readable storage medium storing a plurality of instructions loaded to perform the above-mentioned retraining method of a detection model. An embodiment of the invention provides a retraining device for a detection model, which comprises an input unit and an operation unit. The input unit obtains accuracy information of an object detection model. The operation unit retrains the object detection model, and estimates that the accuracy information is abnormal according to an accuracy supervision model. According to the retraining method and device for the detection model and the non-transitory computer readable storage medium, the retraining of the object detection model can be triggered by the fact that the accuracy monitoring model observes that the object detection model is reduced in efficiency, and manual monitoring is not needed. Drawings FIG. 1 is a schematic diagram of a retraining apparatus for a detection model according to an embodiment of the invention. FIG. 2 is a flowchart of a method for retraining a detection model according to an embodiment of the invention. Fig. 3 is a flowchart of a method for detecting product defects according to an embodiment of the invention. FIG. 4 is a schematic diagram illustrating the efficiency of a supervised object detection model according to an embodiment of the present invention. FIG. 5 is a schematic diagram showing the comparison of the identification type and the real type according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a method for determining performance degradation of an object detection model according to an embodiment of the present invention. FIG. 7 is a schematic diagram of the leak detection rate of a normal test set according to an embodiment of the invention. FIG. 8 is a schematic diagram of the leak detection rate of an anomaly test set according to an embodiment of the present invention. FIG. 9 is a schematic diagram of the false positive rate of a normal test set according to an embodiment of the present invention. FIG. 10 is a schematic diagram of the false positive rate of an anomaly test set according to an embodiment of the present invention. FIG. 11 is a flow chart illustrating implementation of an accuracy supervision model according to an embodiment of the invention. Fig. 12 is a graph showing a change in the false kill rate of an object detection model according to an embodiment of the present invention. Reference numerals illustrate: 10 product(s) 1,1A,1b,1c product photographs 2 Retraining device 3 Input unit 4 Arithmetic unit 5 Non-transitory computer readable storage medium 6,6A ',6b ',6c ' object detection model 7, Precision supervision model 8 Training scheduler G1, g2 frame S t accuracy information S11-S12 step S21-S25 step S31 to S32 steps S41 to S44 step S51-S55 step Interval T1-T5 Detailed Description Referring to fig. 1, a schematic architecture diagram of a retraining apparatus 2 for a detection model according to an embodiment of the invention is shown. The retraining apparatus 2 includes an input unit 3 and an operation unit 4, and a non-transitory computer readable storage medium 5. The computing unit 4 is coupled to the input unit 3 and the non-transitory computer readable storage medium 5. The arithmetic unit 4 acquires a plurality of product photographs 1 via the input unit 3. The product photograph 1 is taken of a plurality of products 10 in the same product type, respectively. The non-transitory computer readable stor