Visibility Restoration: A Systematic Review and Meta-Analysis

Abstract

Image acquisition is a complex process that is affected by a wide variety of internal and environmental factors. Hence, visibility restoration is crucial for many high-level applications in photography and computer vision. This paper provides a systematic review and meta-analysis of visibility restoration algorithms with a focus on those that are pertinent to poor weather conditions. This paper starts with an introduction to optical image formation and then provides a comprehensive description of existing algorithms as well as a comparative evaluation. Subsequently, there is a thorough discussion on current difficulties that are worthy of a scientific effort. Moreover, this paper proposes a general framework for visibility restoration in hazy weather conditions while using haze-relevant features and maximum likelihood estimates. Finally, a discussion on the findings and future developments concludes this paper.

Keywords: haze removal; image defogging; image dehazing; meta-analysis; systematic review; visibility enhancement.

Figure 1

Typical example of digital camera workflow.

Figure 2

PRISMA flow diagram for the systematic review in this study.

Figure 3

Visual illustration of the optical image formation in the atmosphere.

Figure 4

General classification of visibility restoration algorithms.

Figure 5

Simplified block diagrams of: (a) contrast enhancement and (b) polarimetric dehazing approaches in visibility restoration.

Figure 6

Simplified block diagram of dark channel prior-based visibility restoration methods.

Figure 7

Simplified block diagram of image fusion-based visibility restoration methods.

Figure 8

Branching diagram summarizing image-processing-based dehazing algorithms.

Figure 9

Simplified block diagram summarizing regression and regularization-based visibility restoration methods.

Figure 10

Branching diagram summarizing machine-learning-based dehazing algorithms.

Figure 11

Simplified block diagram of convolutional neural network-based visibility restoration algorithms.

Figure 12

Simplified block diagram of generative adversarial network-based visibility restoration algorithms.

Figure 13

Branching diagram summarizing deep-learning-based dehazing algorithms.

Figure 14

Typical procedure for preparing the synthetic training dataset.

Figure 15

Proposed dehazing framework based on a machine learning technique.

Figure 16

Selection of representative hazy patches based on: (a) mean subtracted contrast normalized, (b) sharpness, (c) contrast, (d) entropy, (e) dark channel prior, and (f) saturation features. (g) Selected patches.

Figure 17

Selection of representative haze-free patches based on: (a) mean subtracted contrast normalized, (b) sharpness, (c) contrast, (d) entropy, (e) dark channel prior, and (f) saturation features. (g) Selected patches.

Figure 18

Atmospheric light estimation procedure utilized by: (a) Zhu et al. [52]–red pixels belongs to the top $0.1\%$ brightest pixels, and (b) Park et al. [121]—blue lines represent the quadtree-decomposition process, and the red dot represents the atmospheric light’s estimate.

Figure 19

A qualitative comparison of different dehazing methods on a real hazy image of a train. (a) Hazy image, and results by (b) Tarel and Hautiere [35], (c) He et al. [21], (d) Kim et al. [36], (e) Bui and Kim [50], (f) Zhu et al. [52], (g) Ngo et al. [74], (h) Cai et al. [85], (i) Ren et al. [89], and (j) the proposed framework.

Figure 20

A qualitative comparison of different dehazing methods on a real hazy image of mountains. (a) Hazy image, and results by (b) Tarel and Hautiere [35], (c) He et al. [21], (d) Kim et al. [36], (e) Bui and Kim [50], (f) Zhu et al. [52], (g) Ngo et al. [74], (h) Cai et al. [85], (i) Ren et al. [89], and (j) the proposed framework.

Figure 21

A qualitative comparison of different dehazing methods on a real hazy image of a road scene. (a) Hazy image, and results by (b) Tarel and Hautiere [35], (c) He et al. [21], (d) Kim et al. [36], (e) Bui and Kim [50], (f) Zhu et al. [52], (g) Ngo et al. [74], (h) Cai et al. [85], (i) Ren et al. [89], and (j) the proposed framework.

Figure 22

A qualitative comparison of different dehazing methods on a synthetic hazy image of a road scene. (a) Hazy image, results by (b) Tarel and Hautiere [35], (c) He et al. [21], (d) Kim et al. [36], (e) Bui and Kim [50], (f) Zhu et al. [52], (g) Ngo et al. [74], (h) Cai et al. [85], (i) Ren et al. [89], (j) and the proposed framework, and (k) ground truth.

Figure 23

A qualitative comparison of different dehazing methods on a synthetic hazy image of an indoor scene. (a) Hazy image, results by (b) Tarel and Hautiere [35], (c) He et al. [21], (d) Kim et al. [36], (e) Bui and Kim [50], (f) Zhu et al. [52], (g) Ngo et al. [74], (h) Cai et al. [85], (i) Ren et al. [89], (j) and the proposed framework, and (k) ground truth.