Image data enhancement method and device, medium and equipment

Abstract

The invention discloses an image data enhancement method and device, a medium and equipment, and the method comprises the steps: carrying out the training of a deep learning network according to an image training set, and obtaining a soft label generator, wherein the image training set comprises K types of image sample sets; obtaining a soft label and a category of the first image sample, whereinthe soft label represents the category association degree of the first image sample and the K categories; obtaining category representativeness of the image samples in the K-1 categories of image sample sets according to a soft label generator, wherein the K-1 categories do not contain the category of the first image sample; selecting a second image sample from the K-1 types of image sample sets according to the soft label of the first image sample and the type representativeness of the image sample; and fusing the first image sample and the second image sample to obtain a target image sample.The invention relates to the field of computer vision, and by generating a new image sample with controllable difficulty, the training difficulty degree of a classification network model can be adjusted to improve the generalization ability of the model.

Description

technical field [0001] The present invention relates to the technical field of artificial intelligence, in particular to an image data enhancement method, device, medium and equipment. Background technique [0002] With the development of artificial intelligence, deep learning models have been widely used in various industries and play a very important role in various scenarios. The deep learning network model requires a large amount of sample data for training, and when the amount of sample data is insufficient, the training of the model will cause overfitting. Therefore, a data augmentation method for training data is proposed, that is, operations such as geometric transformation and color transformation are performed on the basis of existing sample data to generate more sample data. [0003] In the field of computer vision, for the image classification network and image training samples, data enhancement methods for mixed samples are also proposed, such as Mixup, Cutout ...

AI Technical Summary

Problems solved by technology

These three methods all improve the generalization of the deep learning classification network by generating difficult examples, but the above methods all randomly select two samples from the training set for fusion to synthesize new samples of difficult examples. Lack of consideration, which makes it difficult to measure and control the difficulty of the synthetic samples for network model training

Taking CutMix as an example, using the fusion of pictures of bicycles and cats, the model can easily distinguish the two categories of bicycles and cats, but if the samples are obtained by fusion of pictures of cats and dogs, it will be difficult for the model to accurately judge the two categories in the picture. Two categories, because dogs are more similar in appearance to cats than bicycles

Images