A hybrid model-based image recognition method, device, and medium

Abstract

The invention discloses an image recognition method based on a mixed model, comprising the following steps: step 1, constructing a feature encoder F from an original image space to a feature space; step 2, constructing a generation model D that fits the probability distribution of the feature space; Step 3, constructing a known class classification model C for K classification of the feature space, where K is the number of known classes; Step 4, jointly training the feature encoder, generation model and classification model to optimize the feature space; Step 5, for The generative model after joint training sets the threshold for identifying unknown classes; step 6, according to steps 1, 2, 3, 4, 5, obtains a K+1 classification hybrid model that can identify images of unknown classes, so as to realize classification and recognition of images to be classified .

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to an image recognition method based on a mixed model. Background technique [0002] Most image recognition methods can only recognize images from known classes, where the known classes represent classes that have been seen during training. However, in real scene applications, the recognition system may face unknown class images from outside the training set category distribution, and these unknown class images will be recognized by the recognition system as one of the known classes, which will lead to problems in practical applications. security issues. Therefore, this requires image recognition methods not only to correctly classify images from known classes, but also to identify images of unknown classes outside the class distribution as unknown classes. [0003] Most of the current related image recognition methods are by fitting the probability distribution of the training imag...

AI Technical Summary

Problems solved by technology

However, this modification is extremely challenging because modifying the known class probability distribution may affect the classification accuracy of the known class

At the same time, the feature space used to fit the probability distribution is based on the learning of known class classifiers, and this classifier focuses on known classes, which may not be accurate for identifying unknown classes.

Images