Image Dedusting with Deep Learning-based Closed-Loop Network

Abstract

Computer Vision Technologies, Including 2D/3D Perceptions, Have Grasped More And More Attentions In the Construction Industry, Which Can Be Employed For Providing Effective Support In Many Industry Tasks, Including Autonomous Excavators, Autonomous Trucks. Etc, And Excavators And Dump Trucks Are The Most Common Assets Used In The Construction Industry. However, Dusts And Mists Always Exist In Many Industry Jobsites, Which Heavily Affects The Applications Of Computer Vision Technologies In the Construction Industry, Such As Segmentation, Detection. Etc. To Solve The Problem, We Propose A Computer-Vision Approach That Utilizes a Deep Convolutional Network Based Closed-Loop Framework To Remove The Influence Of Dust And Mist In Construction Sites. Taking A Image With Dust As Input, The Proposed Framework Can Effectively Remove The Dust Component, Thus Clean Image Can Be Obtained. To Achieve This, We Divide The Image With Dust Into Three Components, Including Clean Image Component, Atmospheric Light Component And Transmission (Dust) Component, And A U-Net Structure That Contains A Encoder And Three Decoders Is Used, Where The Encoder For Effective Feature Extraction, And The Three Decoders For The Recovery Of The Three Image Components, Thus Clean Image, Atmospheric Light And Transmission (Dust) Components Can Be Obtained. To Guarantee That The Network Can Be Convergent, Two Supervisions Are Utilized, One For The Clean Image, And The Sum Of The Three Image Components Should Be the Same With The Input. Experiments Demonstrate The Effectiveness Of The Proposed Approach.