Few-shot Classification Algorithm Based on Latent Spatial Transformation and Spatial Frequency Integration
WANG Xiao
DING Sheng
Abstract：Aiming at the problem of uncertainty caused by a small number of labeled samples in few-shot classification,a se-ries of methods of image operation and a method of combining spatial information and frequency domain information in training are proposed.Firstly,RGB images are converted into YCbCr images in the image preprocessing stage,and then discrete cosine trans-form(DCT)is used to generate frequency domain information to preprocess images and ensure the integrity of information.Second-ly,the feature vectors are preprocessed,and the latent space transformation algorithm is used to modify the class distribution of each class so that it tends to Gaussian distribution,and the optimal transmission algorithm(Sinkhorn)is used to estimate the opti-mal transmission to realize the initial distribution tends to Gaussian distribution and improve the accuracy of classification.Finally,the network is trained with the combination of spatial information and frequency domain information to make full use of the image in-formation to improve the training effect.Experimental results show that in the 5-way 1-shot classification task of Mini-ImageNet,CIFAR-FS and CUB 200 datasets,the classification accuracy reaches 85.24%,88.21%and 94.28%.The classification accuracy of and 5-way 5-shot classification task reaches 90.52%,91.18%and 95.88%.
Keywords：few-shot classificationfrequency domain informationlatent spaceoptimal transmissionGaussian distribution
Publication Date：2025-11-20
Online Publishing Date：2026-01-28(First online date of this platform, not the publication date of the document)
Pages：7( 3240-3246 )