A hybrid oversampling image enhancement method and device, terminal equipment and medium

By constructing a category-reweighted distribution and generating content-aware hybrid samples, the problem of low tail category recognition accuracy of the Vision Transformer model in long-tailed distributed data is solved, thereby improving the tail category recognition performance and enhancing the model's generalization ability.

CN122368682APending Publication Date: 2026-07-10NAT UNIV OF DEFENSE TECH
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NAT UNIV OF DEFENSE TECH
Filing Date
2026-06-05
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing Vision Transformer models have low accuracy and poor generalization ability in identifying tail categories in long-tailed distributed data. Existing solutions such as reweighting, resampling, and data augmentation methods suffer from problems such as accuracy loss or high noise, and cannot effectively improve the robustness of tail category identification.

Method used

By constructing a class-reweighted distribution, mixed samples of foreground and background image patches are generated. Content-aware soft labels are assigned to the mixed samples using the neural activation maps of a pre-trained neural network. The mixture is then trained using a Gaussian perturbation contrastive loss function to generate enhanced samples with physical consistency and semantic accuracy.

Benefits of technology

Without compromising the accuracy of head category recognition, it significantly improves the recognition performance of tail category, enriches the background diversity of tail category, and enhances the model's generalization ability to tail category.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122368682A_ABST
    Figure CN122368682A_ABST
Patent Text Reader

Abstract

This invention provides a method, apparatus, terminal device, and medium for hybrid oversampled image enhancement, including: acquiring an original image dataset; constructing a class-reweighted distribution based on the number of effective samples in each class; sampling foreground image samples from the class-reweighted distribution and background image samples from the long-tail distribution; segmenting the foreground and background image samples into non-overlapping image blocks to obtain a set of foreground image blocks and a set of background image blocks; generating a binary mask for each image block; performing element-wise mixing of the foreground and background image block sets to obtain hybrid image samples; extracting neural activation maps of the foreground and background image samples; and using the neural activation maps to assign content-aware soft labels to the hybrid image samples. This invention can improve the recognition accuracy and generalization robustness of Transformer-based visual recognition models for tail categories in long-tail distribution scenarios.
Need to check novelty before this filing date? Find Prior Art