Non-supervision night image classification method based on feature amplification

Abstract

The invention belongs to the technical field of computer visual identification, and relates to an non-supervised night image classification method based on feature amplification. The method includes training a classification network by adopting a public data set with daytime image classification labels, extracting feature vectors of input images through the classification network, and calculating feature mean values and covariance matrixes of all categories; inputting a night image without a label into the classification network to obtain a pseudo label of the image, and calculating a feature mean value and a covariance matrix of each category of the night image in the feature space according to the pseudo label; carrying out weighted average on the covariance matrixes obtained by daytime and night images of the same type to obtain a final covariance matrix; performing feature sampling according to the feature mean value of each type of night images and the covariance matrix after weighted averaging; and retraining the classification network by using the sampled feature values and the original feature values. According to the invention, the feature distribution of the daytime image with the label is learned, and the night data is amplified on the feature level, so that the unsupervised classification of the night image is realized.

Description

technical field [0001] The invention belongs to the technical field of computer vision recognition, in particular to an unsupervised nighttime image classification method based on feature amplification. Background technique [0002] Image classification is the most classic task in the field of computer vision recognition, and it is also the basis of many other visual problems, which has great practical value and application prospects. Image classification is essentially a pattern classification problem, and its goal is to divide different images into different categories to achieve the smallest classification error. With the success of convolutional neural network (CNN), deep learning has been proven to be an effective solution to image classification problems. [0003] The existing large-scale public datasets involving image classification mainly include ImageNet, COCO, PascalVOC, etc. However, these datasets are basically images collected in daytime environments. Studies...

AI Technical Summary

Problems solved by technology

However, the nighttime images generated by this method do not conform to the real data distribution, and there is also a field gap with the real nighttime data

Images