基于注意力机制的递进式特征提取去雾网络

打开文本图片集

中图分类号：TP391 文献标识码：A 文章编号：2096-4706（2025）22-0030-06

Progressive Feature Extraction Dehazing Network Based on Attention Mechanism

CHENG Xiaoyuan  WANG Bingwen， LI Baoguang，FENG Lei， JIN Nengzhi （1.Gansu Computing Center，Lanzhou，China;2.KeyLaboratoryofAdvanced Computing，Lanzhou ，China）

Abstract：Aiming atthe limitations of existing single image dehazing algorithms in terms ofaccuracyand detail retention， a progressive feature extraction dehazing network based on atention is constructed.Taking AOD-Net as the benchmark framework，thefeatureinteractionpath isreconstructedbypointwiseconvolutionandmulti-dimensionalcollborativeAetion Mechanism，whicheduces thescaleofmodelparametersandimproves thecomputational effciencyThe progresive feature extractionnetworkstructureisdsignedndthemulti-scalefeaturefusionstrategyisusedtoeancetesepaationabilityofthe network tothelong-rangefogconcentrationgradientandhigh-frequencydetails.Themult-scalestructuralsimilarityconstraint andtheadaptiveloss optimization mechanismarefurther integrated tosignificantlyimprove theconsistencyof texture structure andthe balanceofcolordistribution intherestored image.Theexperimentalresultsshowthattheproposednetwork exhibits excellent detail retention ability and visual naturalnessin both synthetic and real fog image scenes.

Keywords： image dehazing; Convolutional Neural Network; Atention Mechanism; multi-scale network

0 引言

恶劣天气下采集到的图像往往受雾霾等的影响，这将降低图像的对比度和细节的可见度，降低图像的质量，从而影响后续的计算机视觉任务[1]。（剩余12013字）