[Paper Review] Classification-driven Single Image Dehazing

Yanting Pei, Yaping Huang+1 more|arXiv (Cornell University)|Nov 21, 2019

Image Enhancement Techniques35 references4 citations

TL;DR

This paper proposes a unified end-to-end CNN architecture that jointly optimizes image dehazing and image classification by integrating a dehazing sub-network (DNet), a classification-driven conditional GAN (CCGAN), and a classification sub-network (CNet). The method significantly improves both dehazing metrics (PSNR, SSIM) and classification accuracy on CUB-200-2011 and Caltech-256, outperforming state-of-the-art methods by leveraging high-level semantic feedback to enhance visual quality and downstream task performance.

ABSTRACT

Most existing dehazing algorithms often use hand-crafted features or Convolutional Neural Networks (CNN)-based methods to generate clear images using pixel-level Mean Square Error (MSE) loss. The generated images generally have better visual appeal, but not always have better performance for high-level vision tasks, e.g. image classification. In this paper, we investigate a new point of view in addressing this problem. Instead of focusing only on achieving good quantitative performance on pixel-based metrics such as Peak Signal to Noise Ratio (PSNR), we also ensure that the dehazed image itself does not degrade the performance of the high-level vision tasks such as image classification. To this end, we present an unified CNN architecture that includes three parts: a dehazing sub-network (DNet), a classification-driven Conditional Generative Adversarial Networks sub-network (CCGAN) and a classification sub-network (CNet) related to image classification, which has better performance both on visual appeal and image classification. We conduct comprehensive experiments on two challenging benchmark datasets for fine-grained and object classification: CUB-200-2011 and Caltech-256. Experimental results demonstrate that the proposed method outperforms many recent state-of-the-art single image dehazing methods in terms of image dehazing metrics and classification accuracy.

Motivation & Objective

• To address the limitation of existing dehazing methods that optimize only for pixel-level metrics (e.g., PSNR, SSIM) without considering high-level vision task performance.
• To investigate whether incorporating classification feedback can improve the visual quality and semantic fidelity of dehazed images.
• To develop an end-to-end deep learning framework that jointly optimizes image dehazing and image classification.
• To demonstrate that dehazed images generated with classification guidance maintain better performance on downstream tasks like image classification.

Proposed method

• The framework consists of three components: a dehazing sub-network (DNet) for image restoration, a classification-driven conditional GAN (CCGAN) to enhance realism and detail, and a classification sub-network (CNet) to provide high-level supervision.
• The CCGAN sub-network is trained to generate dehazed images that are not only perceptually realistic but also optimized for high-level classification performance.
• The entire network is trained end-to-end using a combined loss: a pixel-level reconstruction loss (MSE) for dehazing and a classification loss for CNet.
• The classification sub-network (CNet) is trained on the dehazed output to ensure semantic consistency and to guide DNet toward producing images that preserve discriminative features.
• The classification-driven GAN component uses the CNet’s feature maps as conditional input to the generator, enabling the GAN to produce dehazed images that are both visually plausible and semantically accurate.
• The model is trained on synthetic hazy images but generalizes well to real-world hazy images, as validated in qualitative comparisons.

Experimental results

Research questions

• RQ1 Can integrating high-level classification feedback improve the visual quality and semantic fidelity of dehazed images beyond standard pixel-level metrics?
• RQ2 Does a classification-driven GAN architecture lead to better dehazing performance and improved classification accuracy compared to standard GANs or no GAN supervision?
• RQ3 How does joint optimization of dehazing and classification loss affect the performance of downstream vision tasks on dehazed images?
• RQ4 To what extent does the proposed unified architecture generalize to real-world hazy images despite being trained on synthetic data?
• RQ5 What is the individual contribution of each component (DNet, CCGAN, CNet) to the overall performance gain in dehazing and classification?

Key findings

• On the CUB-200-2011 dataset, the full model (DNet+CCGAN+CNet) achieves a PSNR of 21.2995 and SSIM of 0.8541, outperforming all ablation variants and state-of-the-art methods.
• On CUB-200-2011, the full model achieves 67.7% classification accuracy with ResNet, significantly higher than the 54.4% baseline (DNet only) and 59.7% with DNet+CNet.
• On Caltech-256, the full model achieves 82.6% classification accuracy with ResNet, outperforming the 81.9% baseline (DNet only) and 82.5% with DNet+CCGAN.
• The ablation study confirms that both the classification-driven GAN (CCGAN) and the classification sub-network (CNet) independently improve dehazing metrics and classification accuracy.
• Qualitative results on real hazy images show that the proposed method produces clearer, less distorted outputs than state-of-the-art methods, with minimal color distortion and effective haze removal.
• The model generalizes well to real-world hazy images despite being trained on synthetic data, demonstrating robustness and practical applicability.