Loss Functions for Image Restoration With Neural Networks

By Hang Zhao, Orazio Gallo, Iuri Frosio, Jan Kautz. IEEE Transactions on Computational Imaging, 2017

2,469 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition, Media Technology

View Paper: https://www.semanticscholar.org/paper/78357f1920fa83e3183acf15d2449e49b9ebbd7d

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Media Technology

Citations in the corpus

Interactive reconstruction of Monte Carlo image sequences using a recurrent denoising autoencoder (2017, positive): “However, it is also known that using L1 loss instead of L2 can reduce the splotchy artifacts from reconstructed images [Zhao et al. 2016].”
Learned Image Downscaling for Upscaling Using Content Adaptive Resampler (2019, background): “The most common loss function generally defaults to the SR network training is the mean squared error (MSE) or the L2 norm loss, but it tends to lead to poor image quality as perceived by human observers [42].”
Deep Fluids: A Generative Network for Parameterized Fluid Simulations (2018, background): “ess is repeated by random batches until the network minimizes a loss function over all the training data. While previous works have proposed loss functions for natural images, e.g., Lp norms, MS-SSIM [ZGFK17], and perceptual losses [JAFF16, LTH17], accurate reconstructions of velocity ﬁelds have not”
Convolutional Neural Networks for Inverse Problems in Imaging: A Review (2017, methodology): “This formulation is used in most of the works we surveyed, [6], [7], [11], [12], [18], [19], [23], [25], [26], despite the fact that several raise questions about whether it is the best choice [25], [34].”
Face identity disentanglement via latent space mapping (2020, methodology): “For this end, we adopt the "mix" loss suggested in Zhao et al. [2016], and use a weighted sum of L1 loss and MS-SSIM loss: Lmix = α (1 −MS-SSIM(Iattr , Iout )) + (1 − α) ∥Iattr − Iout ∥1 (6)”
Emerging From Water: Underwater Image Color Correction Based on Weakly Supervised Color Transfer (2017, methodology): “To preserve the content and structure of the input image, SSIM loss [33] is used.”
Lightweight Pyramid Networks for Image Deraining (2018, methodology): “Following [44], we choose 1 and Structural SIMilarity index (SSIM) [45] as our loss functions.”
Attention-Guided Network for Ghost-Free High Dynamic Range Imaging (2019, result): “Quantitative comparisons of the results are shown in Table 2, which implies that the l1 loss is more powerful for preserving details as discussed in [45].”
Deep reverse tone mapping (2017, methodology): “We use L1 distance instead of L2 distance because it can reduce blurring in the results [Zhao et al. 2017].”
Digging Into Self-Supervised Monocular Depth Estimation (2018, methodology): “Following [21] we use bilinear sampling to sample the source images, which is locally sub-differentiable, and we follow [75, 15] in using L1 and SSIM [64] to make our photometric error function pe, i.”
CT Super-Resolution GAN Constrained by the Identical, Residual, and Cycle Learning Ensemble (GAN-CIRCLE) (2018, methodology): “However, the output optimized by the L 2 norm may suffer from over-smoothing as discussed in [50], [51], since the L 2 distance means to maximizing the peak signal-to-noise rate (PSNR) [30].”
A Sparse-View CT Reconstruction Method Based on Combination of DenseNet and Deconvolution (2018, background): “Multi-scale structure similarity (MS-SSIM) [53] has been proposed to preserve the global structure similarity between images.”
fastMRI: An Open Dataset and Benchmarks for Accelerated MRI (2018, methodology): “It is sometimes advantageous to use the L1 loss for training machine learning models on computer vision tasks, even when evaluation is performed under L2 norm losses such as MSE [59].”
Remote Sensing Image Fusion Based on Two-Stream Fusion Network (2017, background): “Several studies have suggested that (cid:96) 1 loss is a better choice when performing image restoration tasks [46, 55], thus in this work, we also prefer (cid:96) 1 as loss function to train our network.”
A Review on Deep Learning Techniques for Video Prediction (2020, background): “Besides video prediction, several studies analyzed the impact of different loss functions in image restoration [39], classification [40], camera pose regression [41] and structured prediction [42], among others.”
UIEC^2-Net: CNN-based Underwater Image Enhancement Using Two Color Space (2021, methodology): “The L1 losss presented in Equation 4: SSIM loss( L SSIM ): We use the SSIM loss [51] in our ob-jective function to impose the structure and texture similarity on the predicted image.”
The Sound of Pixels (2018, methodology): “In this case, per-pixel L1 loss [47] is used for training.”
Deep Generative Adversarial Networks for Compressed Sensing Automates MRI (2017, background): “Also, (cid:96) 1 -net performs better than 2 -net, which was also already reported in a different setting [21], but still not as sharp as GANCS which leverages both (cid:96) 1 -net and GAN.”
SuperYOLO: Super Resolution Assisted Object Detection in Multimodal Remote Sensing Imagery (2022, methodology): “The L1 loss (rather than L2 loss) [44] is used to calculate the SR construction loss L s between the input image X and SR result S , to which the expression is written as The detection loss involves three components [19]: loss of judging whether there is an object L obj , loss of object location L…”