Learning Spatiotemporal Features for Single Image Stripe Noise Removal

Juntao Guan,Ai Xiong,Rui Lai

doi:10.1109/access.2019.2944239

Juntao Guan, Ai Xiong + Show 1 more

Open Access

https://doi.org/10.1109/access.2019.2944239

Copy DOI

Abstract

The stripe noise caused by nonuniform response of different detectors limits the sensitivity of the infrared imaging system and reduces the image quality. Existing destriping methods still struggle to remove the stripe noise as well as recover the image details, which restricts the application of the infrared focal plane array (IRFPA) imager. In this paper, an innovative destriping method through the perspective of spatiotemporal feature modeling is proposed, which excavates the intrinsic spatial characteristics of stripe noise as well as the redundant temporal information between the adjacent frames to estimate the stripe component more precisely. Moreover, the bidirectional fusion strategy that further strengthens the long-time correlation is introduced to separate the scene details from stripe noise more thoroughly. Experimental results show that the proposed model outperforms existing classical destriping methods on both simulated images and real data.

Highlights

Infrared imaging system plays an import role in a wide range of fields, such as remote sensing, military and industry inspection [1]
In order to validate the effectiveness of the proposed model, the classical image benchmark Set14 [40] is adopted as the test dataset, and it is not included in the training pairs
EXPERRIMENTAL RESULTS AND ANALYSIS we will carry out the explorative experiments to reveal the properties of our presented model

Summary

INTRODUCTION

Infrared imaging system plays an import role in a wide range of fields, such as remote sensing, military and industry inspection [1]. Inspired by the roaring success of deep learning in image restoration and forecasting tasks, deep neural networks are gradually introduced to remove the stripe from complex scenes [9]–[11] He et al employ the convolutional neural network (CNN) with residual connection to estimate the stripe component directly [12]. Chang et al propose a multiscale residual deep convolutional neural network to remove the stripe noise [14] These methods ignore the self-characteristics of stripe component, which causes the detail loss in the destriping results [15], [16]. The characteristics of the stripe noise produced by the vertical adjacent detectors is time related, which shows special temporal dependency for each stripe In this work, both of the spatial characteristic and temporal dependency will be utilized to remove the stripe noise as well as preserve the scene details efficiently

RECURRENT NEURAL NETWORK FOR SEQUENCE MODELING

GATED CONVOLUTION RECURRENT UNIT

BIDIRECTIONAL GCRU

EXPERRIMENTAL RESULTS AND ANALYSIS

CONCLUSION