Ten Deep Learning Based Image Dehazing Algorithms: Principles, Implementations, and Comparisons

Image dehazing aims to restore clear visibility in images degraded by atmospheric scattering, and it has become a popular research topic in computer vision and artificial intelligence. The following sections summarize ten representative deep‑learning based dehazing approaches, each accompanied by a brief description, key advantages, limitations, and references to the original papers and source code.

1. DehazeNet

DehazeNet is a convolutional neural network that estimates atmospheric light and transmission map in an end‑to‑end fashion. It consists of feature extraction, multi‑scale mapping, local extremum, and non‑linear regression units built from three convolutional layers, one pooling layer, a Maxout unit and a BReLU activation. The network parameters are shown in the accompanying diagram.

Training steps include synthetic dataset generation, network architecture design, loss function (perceptual + smoothness), iterative parameter updates, and testing on unseen hazy images. Advantages: good dehazing quality, adaptive parameter adjustment, enhanced depth perception. Drawbacks: high computational cost, poor performance on low‑contrast images, sensitivity to noise.

2. MSCNN (Multi‑scale CNN)

MSCNN combines multi‑scale convolutional features with holistic edge information. The pipeline consists of dataset preparation, multi‑scale CNN training, holistic edge generation, and dehazing using the learned mapping. The method uses residual and perceptual losses.

Advantages: high‑quality results, multi‑scale processing, end‑to‑end training, good robustness. Disadvantages: high computational complexity, sensitivity to input parameters, some detail loss.

3. AOD‑Net

AOD‑Net (All‑in‑One Dehazing Network) redesigns the atmospheric scattering model and directly generates a clean image via a lightweight CNN, without separately estimating transmission or atmospheric light. It consists of a K‑estimation module (five convolutional layers with multi‑scale concatenations) and a clean‑image generation module (element‑wise multiplication and addition).

Key code snippet:

import torch
import torch.nn as nn
import math
class dehaze_net(nn.Module):
    def __init__(self):
        super(dehaze_net, self).__init__()
        self.relu = nn.ReLU(inplace=True)
        # each conv layer uses only three kernels
        self.e_conv1 = nn.Conv2d(3,3,1,1,0,bias=True)
        self.e_conv2 = nn.Conv2d(3,3,3,1,1,bias=True)
        self.e_conv3 = nn.Conv2d(6,3,5,1,2,bias=True)  # concat1+2 -> 6
        self.e_conv4 = nn.Conv2d(6,3,7,1,3,bias=True)  # concat2+3 -> 6
        self.e_conv5 = nn.Conv2d(12,3,3,1,1,bias=True) # concat all ->12
    def forward(self, x):
        source = []
        source.append(x)
        # K‑estimation
        x1 = self.relu(self.e_conv1(x))
        x2 = self.relu(self.e_conv2(x1))
        concat1 = torch.cat((x1,x2), 1)
        x3 = self.relu(self.e_conv3(concat1))
        concat2 = torch.cat((x2, x3), 1)
        x4 = self.relu(self.e_conv4(concat2))
        concat3 = torch.cat((x1,x2,x3,x4),1)
        x5 = self.relu(self.e_conv5(concat3))
        clean_image = self.relu((x5 * x) - x5 + 1)
        return clean_image

Advantages: lightweight, end‑to‑end, can be embedded in downstream tasks. Limitations: modest performance on heavy haze, requires careful training.

4. NLD (Non‑local Dehazing)

NLD leverages non‑local similarity across the whole image to estimate transmission and restore scene radiance. It first estimates the haze model, computes the dark channel prior, estimates haze density, calculates non‑local similarity weights, and finally removes haze via weighted averaging.

Advantages: high‑quality results, fully data‑driven, no extra inputs required. Drawbacks: high computational cost, noise sensitivity, parameter tuning challenges.

5. SSLD (Semi‑Supervised Image Dehazing)

SSLD combines labeled synthetic pairs with unlabeled real hazy images. It uses a supervised branch trained with MSE, perceptual, and adversarial losses, and an unsupervised branch trained with dark‑channel and total‑variation losses. The architecture consists of two identical weight‑sharing branches.

Pros: high‑quality dehazing, flexible semi‑supervised learning, strong extensibility. Cons: requires some labeled data, sensitive to haze density variations, needs careful parameter adjustment.

6. EPDN (Enhanced Pix2Pix Dehazing Network)

EPDN builds on the pix2pix GAN framework with multi‑resolution generators (global G1 and local G2), multi‑scale discriminators (D1, D2), and enhancement blocks that use multi‑scale pyramids and 1×1 convolutions. It employs perceptual, content, and adversarial losses.

Strengths: high‑quality results, multi‑scale processing, enhanced detail recovery. Weaknesses: high computational demand, sensitive to training data quality.

7. DAD (Domain Adaptation for Image Dehazing)

DAD addresses domain shift between synthetic and real hazy images by aligning feature distributions using Maximum Mean Discrepancy (MMD) loss and Adaptive Batch Normalization. It includes a translation module and two dehazing modules for source and target domains.

Benefits: improved cross‑domain performance, unsupervised adaptation. Limitations: reliance on sufficient target data, sensitivity to illumination changes.

8. PSD (Principled Synthetic‑to‑Real Dehazing)

PSD integrates physical priors (e.g., atmospheric scattering equations) with a deep network trained on synthetic hazy images. The network predicts transmission maps, which are refined using physical constraints before reconstructing the clean image.

Pros: principled guidance, strong visibility improvement. Cons: heavy training data dependence, high computational cost, illumination sensitivity.

9. MSBDN (Multi‑scale Boosted Dehazing Network with Dense Feature Fusion)

MSBDN employs a U‑Net‑like encoder‑decoder with dense feature fusion and boosting modules across multiple scales. It combines coarse and fine features via skip connections and uses a multi‑scale discriminator for adversarial training.

Advantages: multi‑scale processing, boosted dehazing, dense fusion, high performance. Drawbacks: high computational load, less effective on extreme haze, sensitive to input quality.

10. GFN (Gated Fusion Network)

GFN fuses features from different depths using a gating mechanism. It contains a feature extractor, haze perception module, layer‑wise fusion module, and dehaze restoration module. Dilated convolutions enlarge receptive fields, and three confidence maps weight the contributions of three enhanced versions of the input.

Pros: good dehazing quality, low computational cost, residual learning. Cons: requires large training datasets, sensitive to image quality, inference latency.

Comparison Results

After presenting the principles of the ten algorithms, the article shows side‑by‑side visual comparisons of their dehazing results on benchmark images, illustrating the strengths and weaknesses of each method.

Overall, the survey provides a comprehensive reference for researchers and practitioners interested in state‑of‑the‑art image dehazing techniques.