伴隨深度學習在圖像修復任務中的發展，諸多方法透過注意力機制來加強模型對
圖像淺層特徵的處理，進而在合成數據集中達到理想修復效果。由於現今大多數圖像去霧的方法皆以訓練在合成數據集為主，因此將模型運用在真實霧霾情況下的去霧效果並非理想，然而在存有的真實數據集中可訓練資料量不多，因此在有限數量中訓練模型達到理想去霧效果已成為一個挑戰。在有限資料情況下，透過多尺度的模型設計提升圖像品質是廣為人知的手法。

多尺度卷積網路藉由提取不同尺寸的圖像之整體特徵及細節資訊來達到模型之效
能提升，而鑑於深度學習近期的發展下，多尺度架構設計能增強圖像處理之效能已被證實。然而，霧霾的型態千變萬化且無可預知，是故只透過多尺度的特徵提取是不足以應變多樣化的霧霾情況。

在本篇論文中提出基於多尺度關注之金字塔卷積的去霧架構。首先，為了增加其
延展性，因此增加不同尺度大小之金字塔卷積（Convolution）架構；再者，為了達到有效去霧，透過注意力機制使模型關注圖片的高頻資訊。實驗結果顯示本篇論文相較於過去方法的優勢。

With the development of deep learning in image restoration tasks, many deep learning-based methods enhance the abilities of models for extracting the effective features of images through attention mechanisms, achieving ideal dehazing performance. Since existing deep learning approaches of image dehazing tasks are mainly trained on synthetic datasets, the domain gap between real-world haze and synthetic haze makes the trained models have performance drop obviously. For the existing limited real-world datasets, it has been a challenge to train models with limited data. With the limitation of available real-world data, multiscale architecture is a well-known technique to improve model performance.

Multiscale convolutional networks achieve performance enhancement by extracting global information and local details from different scale features, and it has been proven that the multiscale architecture can improve the performance in image restoration tasks. However, haze is variable and unpredictable, it is not sufficient to handle diverse haze conditions with only consider multiscale extracted features.

We propose a Multiscale Attention-guided Pyramid Network for single image dehazing tasks in our method. In order to increase the extensibility of the model, we apply a pyramid convolutional block. Apart from this, the attention mechanisms make the model pay attention to the high-frequency region in order to dehaze effectively. The extensive evaluation results show the advantages of our proposed method compared to previous works.

摘要 ii
Abstract iii
1 Introduction 7
2 Related Work 9
2.1 Handcrafted prior-based methods 9
2.2 Learning-based methods 10
3 Methodology 11
3.1 Overview 11
3.2 Multiscale Attention-guide block (MA block) 12
3.2.1 Basic Block Structure 13
3.3 Pyramid Convolutional block (PC block) 14
3.4 Loss Functions 15
3.4.1 Reconstruction loss 15
3.4.2 Perceptual loss 16
3.4.3 Laplacian loss (Edge loss) 16
4 Experiments 17
4.1 Datasets 17
4.1.1 Homogeneous Haze 17
4.1.2 Non-Homogeneous Haze 17
4.2 Implementation Details 18
4.3 Performance Evaluation 18
4.4 Qualitative Visualization 19
4.5 Ablation Study 22
4.5.1 Architecture ablation 22
4.5.2 Loss ablation 23
5 Conclusion 25
Reference 26

[1]E.J. McCartney. “Optics of the atmosphere”. In:scattering by molecules and parti-cles, New York, John Wiley and Sons, Inc.,1976, p. 421.
[2]Avidan S. et al. Berman D. “Non-local image dehazing”. In:Proceedings of the IEEEconference on computer vision and pattern recognition,1974-1682.
[3]Sun C. Zhao Y. Jiang Y. and L. Yang. “Image dehazing using adaptive bi-channelpriors on superpixels”. In:Computer Vision and Image Understanding, 165:17–32.
[4]Sun J. Tang X. He K. “Single image haze removal using dark channel prior”.In:IEEE transactions on pattern analysis and machine intelligence 33(12). 2010,pp. 2341–2353.
[5]Ma Y. Shi Z.-Chen J. Liu X. “Griddehazenet: Attention-based multi-scale networkfor image dehazing”. In:Proceedings of the IEEE International Conference on Com-puter Vision. 2019, pp. 7314–7323.[6]Xu X. Jia K.-Qing C. Cai B. and D. Tao. “Dehazenet: An end-to-end system forsingle image haze removal”. In:IEEE Transactions on Image Processing 25(11).2016, pp. 5187–5198.
[7]R. Fattal. “Single image dehazing”. In:ACM transactions on graphics (TOG)27(3):72. 2008, pp. 1–9.
[8]Sun J. Tang X. He K. “Single image haze removal using dark channel prior”.In:IEEE transactions on pattern analysis and machine intelligence 33(12). 2010,pp. 2341–2353.
[9]R. Fattal. “Dehazing using color-lines”. In:ACM transactions on graphics (TOG)34(1). 2014, pp. 1–14.
[10]Liu S. Zhang H.-Pan J. Cao X. Ren W. and M.H. Yang. “Single image dehazing viamulti-scale convolutional neural networks”. In:European Conference on ComputerVision. 2016, pp. 154–169.
[11]Wang Z. Bai Y.-Xie X. Qin X. and H. Jia. “FFA-Net: Feature Fusion AttentionNetwork for Single Image Dehazing”. In:AAAI. 2020, pp. 11908–11915.
[12]He Zhang and Vishal M Patel. “Densely Connected Pyramid Dehazing Network”.In:CVPR. 2018.
[13]Ajay Bhave Ayush Singh and Dilip K. Prasad. “Single image dehazing for a varietyof haze scenarios using back projected pyramid network”. In:ECCV Workshop.2020.
[14]Fischer P. Brox T. Ronneberger O. “U-Net: convolutional networks for biomedicalimage segmentation”. In:International Conference on Medical Image Computingand Computer Assisted Intervention. 2015, pp. 234–241.
[15]Ancuti C. O. Timofte-R. Ancuti C. “Ntire 2018 challenge on image dehazing: Meth-ods and results”. In:Proceedings of the IEEE Conference on Computer Vision andPattern Recognition Workshops. 2018, pp. 891–901.
[16]Gu S. Timofte R.-Zhang L. Cai J. “Ntire 2019 challenge on real image super-resolution: Methods and results”. In:Proceedings of the IEEE Conference on Com-puter Vision and Pattern Recognition Workshops. 2019, pp. 0–0.
[17]Ancuti C.O. Vasluianu F.A.-Timofte R. Ancuti C. “Ntire 2020 challenge on nonho-mogeneous dehazing”. In:Proceedings of the IEEE/CVF Conference on ComputerVision and Pattern Recognition Workshops. 2020, pp. 490–491.
[18]Suganuma M. Sun Z. Liu X. and T. Okatani. “Dual Residual Networks Leverag-ing the Potential of PairedOperations for Image Restoration”. In:Conference onComputer Vision and Pattern Recognition. 2019b, pp. 7007–7016.
[19]Huang Ziling Tsai Chung-Chi Deng Qili and Chia-Wen Lin. “HardGAN: A Haze-Aware Representation Distillation GAN for Single Image Dehazing”. In:EuropeanConference on Computer Vision. Springer. 2020, pp. 722–738.