Abstract: To address underexposure and overexposure in images,a multi-level information fusion exposure correction network based on the Laplacian pyramid structure was developed.Each network level adopted a U-Net-like encoder-decoder architecture in its correction module.A multi-scale convolutional encoder based on ConvNeXt-tiny was designed as the primary feature extraction unit to enhance feature extraction ability while reducing the model's parameter count.To tackle the issue of checkerboard artifacts arising during image up-sampling,a dual-path up-sampling module combining bilinear interpolation and sub-pixel convolution was proposed.The network demonstrated effective results in both quantitative and qualitative validations on a large-scale exposure correction dataset.Dowel positioning experiments showed significant improvements in feature repeatability,positioning accuracy,and stability at varying contrast thresholds when the network was applied to image enhancement.

Key words: information fusion, exposure correction, multi-scale convolutional encoder, dual-path up-sampling

摘要： 针对图像曝光不足和曝光过度的问题,设计了基于拉普拉斯金字塔结构的多层级信息融合曝光矫正网络,该网络每个层级的矫正模块采用类U-Net的编码器解码器架构。为了提高矫正模块的特征提取能力并减少模型参数量,设计了基于ConvNeXt-tiny的多尺度卷积编码器作为基本特征提取单元。针对图像上采样过程中可能出现的棋盘格伪影问题,提出一种基于双线性插值和亚像素卷积的双路上采样模块。通过定量和定性验证,在大规模曝光矫正数据集上均取得较好的结果。定位销定位实验显示,在不同对比度阈值下,应用该网络进行图像增强显著提升了特征可重复性、定位精度和稳定性。

关键词: 信息融合, 曝光矫正, 多尺度卷积编码器, 双路上采样