Image processing method based on FoldNet network

By introducing fold convolution and adaptive bidirectional hybrid modules into the FoldNet network, the shortcomings of CNN in capturing global information are solved, and the global representation ability and performance of CNN are improved without increasing computational complexity.

CN119723147APending Publication Date: 2025-03-28NANJING UNIV OF POSTS & TELECOMM
0 Cites 0 Cited by

Patent Information

Application Number
CN202411497309.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-25
Publication Date
2025-03-28

AI Technical Summary

Technical Problem

Existing convolutional neural networks (CNNs) have shortcomings in capturing global information, making it difficult to achieve similar performance to Vision Transformer (ViT), and the high computational complexity of ViT limits its application on edge devices.

Method used

An image processing method based on FoldNet network is designed, including a folded convolution module for global information and an adaptive bidirectional hybrid module. Global features are extracted through feature mapping folding, spatial grouping convolution and global information encoding, and global features are selected and fused through joint representation generation, feature refinement and fusion.

Benefits of technology

It realizes that the global representation capability of CNN is improved without increasing computational complexity, and achieves similar performance to ViT, while reducing model size and computing resource requirements, which are suitable for edge devices.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention provides an image processing method based on a FoldNet network. The method comprises the following steps: providing a global information-oriented module, namely folding convolution; designing an adaptive bidirectional hybrid module, and respectively extracting local and global features by using a MobileNetV2 module and folding convolution; a series of FoldNet networks are established based on the adaptive bidirectional hybrid module; using the network structure to train data to obtain a training model, using the verification set to perform verification and parameter adjustment, and selecting an optimal model; and testing the selected optimal model by using the test set. According to the method, the effectiveness of the FoldNet in image classification is verified by comparing a large number of experiments with an advanced model, good balance can be achieved between calculation complexity and accuracy, and meanwhile the method can be transferred and applied to many downstream tasks such as detection, segmentation and generation.
Need to check novelty before this filing date? Find Prior Art