级联残差学习的红外图像非均匀性校正方法

Cascade residual learning method for infrared image nonuniformity correction

针对现有场景自适应非均匀性校正方法存在的图像过平滑和非均匀性残留问题,提出了一种基于级联残差学习的非均匀性校正方法。该方法将多尺度特征提取单元所获取的特征进行融合,并运用残差学习策略解决深度神经网络的过拟合问题。实验结果表明,该方法在平均峰值信噪比上较传统的场景自适应校正方法有近5dB的提升,主观视觉效果也更加清晰锐利。

Traditional scene adaptive nonuniformity correction methods generally suffer from the over smooth and residual nonuniformity in the corrected results.In view of this,a cascade residual learning based nonuniformity correction method is presented.This method uses the multiscale feature extraction unit to fuse the extracted features and employs the residual learning strategy to deal with the overfitting problem.Experimental results validate that the proposed method yields nearly 5dB improvement in the average peak signal-to-noise ratio(PSNR)as compared to the traditional scene adaptive correction methods.Moreover,its visual effects are clearer and sharper.