Fine-grained classification denoising training method based on prediction confidence
The invention discloses a fine-grained classification denoising training method based on prediction confidence, and the method comprises the steps: S1, firstly enabling all training samples to participate in preheating training, and recording the recent several prediction results of each sample as a...
Saved in:
Main Authors | , , , , |
---|---|
Format | Patent |
Language | Chinese English |
Published |
31.01.2023
|
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | The invention discloses a fine-grained classification denoising training method based on prediction confidence, and the method comprises the steps: S1, firstly enabling all training samples to participate in preheating training, and recording the recent several prediction results of each sample as a historical prediction set; s2, generating normalized prediction confidence of each sample through a histogram generated by a historical prediction set; and S3, balancing weights of sample labels and sample prediction by adopting normalized prediction confidence, and dynamically correcting a loss value. According to the method, the dynamic loss replaces the common cross entropy loss and is used for distinguishing the distributed external noise and other samples, so that the distributed external noise can be better removed; when the model is trained on a noisy data set, de-noising training is carried out through loss correction and a global sample selection strategy in a framework, and the classification precision o |
---|---|
Bibliography: | Application Number: CN202211452486 |