摘要
当前,为实现图像全局建模的目的,基于多层感知机(multi-layer perceptron,MLP)的模型通常需要将图像上的像素进行平铺,之后实施一个自注意力机制或“混合”增强方案以获得图像的长范围依赖。然而,这些方法通常消耗大量的计算资源来弥补图像重建丢失的空间拓扑信息。特别是对于超高清图像去雾任务,大量堆积MLP的模型在资源受限的设备上执行一张超高清带雾图像时会出现内存溢出的问题。为了解决这个问题,本文提出了一种可以在单个GPU上对分辨率为4 k的图像进行实时去雾(110 f/s)的模型,该模型的建模过程中保持了图像空间结构信息,同时具有低计算复杂度的优点。
Current multilayer perceptron(MLP)-based models usually require flattening pixels on an image and subsequently enforce a self-attention mechanism or“Mix”enhancement scheme to achieve global modeling of images and obtain long-range dependence of the image.However,these approaches generally consume considerable computing resources to bridge the loss of spatial topological information in image reconstruction.Particularly for UHD image dehazing tasks,numerous stacked MLP models suffer from memory overflow when running a UHD-hazed image on a resource-constrained device.A novel model for real-time dehazing of 4 K images on a single GPU(110 fps)is proposed here to address this issue.This model is advantageous because it maintains spatial information of the raw image and has low computational complexity.
作者
郑卓然
魏绎汶
贾修一
ZHENG Zhuoran;WEI Yiwen;JIA Xiuyi(School of Computer Science and Engineering,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《智能系统学报》
CSCD
北大核心
2024年第1期89-96,共8页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(62176123)。
关键词
图像去雾
超高清图像
多层感知机
空间拓扑信息
局部特征提取
全局特征提取
深度学习
实时去雾
image dehazing
UHD image
multilayer perceptron
spatial topology information
local feature extraction
global feature extraction
deep learning
real-time image dehazing