Abstract
Single image dehazing has always been a challenging problem in the field of computer vision. Traditional image defogging methods use manual features. With the development of artificial intelligence, the defogging method based on deep learning has developed rapidly. In this paper, we propose a novel image defogging approach called NIN-DehazeNet for single image. This method estimates the transmission map by NIN-DehazeNet combining Network-in-Network with MSCNN (Single Image Dehazing via Multi-Scale Convolutional Neural Networks). In the test stage, we estimate the transmission map of the input hazy image based on the trained model, and then generate the dehazed image using the estimated atmospheric light and computed transmission map. Extensive experiments have shown that the proposed algorithm outperforms traditional methods.
Introduction
Image dehazing—recovering a clear scene from a single hazy, foggy, or smoky image—is a notoriously difficult, ill-posed problem in computer vision. Traditional methods, such as the Dark Channel Prior (DCP), rely on hand-crafted statistical features to estimate atmospheric scattering. While effective in many cases, these manual features often fail in complex scenarios like dense fog or grayscale images, often leading to artifacts or oversaturated colors.
Recent deep learning methods have improved upon this by learning to extract fog-related features end-to-end. However, existing architectures often use standard linear convolutional layers or maxout units, which struggle to model the highly complex, non-linear distributions of latent fog features. NIN-DehazeNet addresses this by designing a network capable of more sophisticated local feature extraction.
Proposed Method: NIN-DehazeNet
The core innovation of NIN-DehazeNet is the integration of the Network-in-Network (NIN) architecture into the Multi-Scale Convolutional Neural Network (MSCNN) framework. Instead of using standard linear convolutional filters, the model employs mlpconv (multilayer perceptron convolution) layers.
An mlpconv layer acts as a "micro-network"—a universal function approximator—that slides over the input image. It maps local patches to output feature vectors using a multilayer perceptron with rectified linear unit (ReLU) activations. This allows for complex, learnable interactions of cross-channel information, enabling the model to capture intricate fog-related features that linear convolutions might miss.
Multi-Scale Architecture
The overall architecture follows a coarse-to-fine refinement strategy:
- Coarse-scale Network: Predicts a holistic, global transmission map of the scene to capture the overall "thickness" of the fog.
- Fine-scale Network: Refines the global map to capture detailed local depth variations and object boundaries.
The final transmission map is used alongside an estimate of global atmospheric light to mathematically reverse the haze using a standard atmospheric scattering model.
Experimental Results
The model was trained on 10,500 synthetic hazy images randomly sampled from the RESIDE (REalistic Single Image DEhazing) dataset. Quantitative evaluation was performed on a custom test set of 100 hazy images derived from video frames.
NIN-DehazeNet was compared against traditional methods (DCP) and state-of-the-art deep learning models including DehazeNet, MSCNN, AOD-Net, and DCPDN. The proposed model achieved superior quantitative results:
- PSNR: 18.48 dB (outperforming MSCNN's 18.22 and DCPDN's 18.46).
- SSIM: 0.862 (outperforming MSCNN's 0.833 and DCPDN's 0.853).
In terms of efficiency, the model achieved an average running time of 0.039 seconds per image using GPU acceleration. This performance is well within the requirements for real-time video processing (~25 frames per second).
Conclusion
NIN-DehazeNet demonstrates that non-linear feature extraction via mlpconv layers significantly improves transmission map estimation for single image dehazing. By combining local micro-networks with a multi-scale refinement architecture, the model achieves state-of-the-art results while maintaining real-time viability. Future work will explore the integration of Generative Adversarial Networks (GANs) to further improve performance on real-world, non-homogeneous fog distributions.
Summary updated