A shadow removal method based on dynamic alignment and illumination-aware convolution

By improving the Unet network structure, using dynamic aligned convolution and bilinear interpolation upsampling, and combining it with a dynamic weight module, the problem of large number of network parameters in existing shadow removal methods is solved, achieving efficient shadow removal and improved robustness.

CN115937030BActive Publication Date: 2026-07-17HANGZHOU DIANZI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HANGZHOU DIANZI UNIV
Filing Date
2022-12-12
Publication Date
2026-07-17

AI Technical Summary

Technical Problem

Existing shadow removal methods suffer from problems such as large network size and high training difficulty, especially deep learning methods based on generative adversarial networks and classical lighting models, which have huge parameter numbers and insufficient robustness.

Method used

An improved method based on the Unet network structure is adopted. By replacing the transposed convolution operation in the decoder part, using Dynamic Aligned Convolution (DAIC) and bilinear interpolation upsampling, and combining it with Dynamic Weight Module (DWM), the network structure is simplified, the number of parameters is reduced, and illumination changes are handled through feature alignment and dynamic convolution weights.

Benefits of technology

While reducing the number of network parameters by 62%, the final error remained essentially unchanged, improving the effect and robustness of shadow removal and effectively addressing changes in local lighting and object material.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115937030B_ABST
    Figure CN115937030B_ABST
Patent Text Reader

Abstract

The application discloses a shadow removal method based on dynamic alignment and illumination perception convolution. In order to reduce the unet network parameter quantity, the method removes the transpose convolution operation in the unet network decoder part, and uses the DAIC and the bilinear interpolation up-sampling method designed in the present application in the unet structure decoder part, so that the network is simplified, the network parameter quantity is reduced to 62% of the original, and the final error is basically unchanged. The FAM is designed for aligning features, can solve the offset features generated through a series of down-sampling, and the features after alignment are more helpful for the image repair work. The DWM is designed for generating dynamic convolution parameters, can generate corresponding convolution weights according to different light changes, and better performs the shadow removal operation on the shadow samples.
Need to check novelty before this filing date? Find Prior Art