DEA-Net: Single image dehazing based on detail-enhanced convolution and content-guided attention (2301.04805v1)

Abstract: Single image dehazing is a challenging ill-posed problem which estimates latent haze-free images from observed hazy images. Some existing deep learning based methods are devoted to improving the model performance via increasing the depth or width of convolution. The learning ability of convolutional neural network (CNN) structure is still under-explored. In this paper, a detail-enhanced attention block (DEAB) consisting of the detail-enhanced convolution (DEConv) and the content-guided attention (CGA) is proposed to boost the feature learning for improving the dehazing performance. Specifically, the DEConv integrates prior information into normal convolution layer to enhance the representation and generalization capacity. Then by using the re-parameterization technique, DEConv is equivalently converted into a vanilla convolution with NO extra parameters and computational cost. By assigning unique spatial importance map (SIM) to every channel, CGA can attend more useful information encoded in features. In addition, a CGA-based mixup fusion scheme is presented to effectively fuse the features and aid the gradient flow. By combining above mentioned components, we propose our detail-enhanced attention network (DEA-Net) for recovering high-quality haze-free images. Extensive experimental results demonstrate the effectiveness of our DEA-Net, outperforming the state-of-the-art (SOTA) methods by boosting the PSNR index over 41 dB with only 3.653 M parameters. The source code of our DEA-Net will be made available at https://github.com/cecret3350/DEA-Net.

Assessment of "DEA-Net: Single Image Dehazing Based on Detail-Enhanced Convolution and Content-Guided Attention"

The paper introduces DEA-Net, a novel approach to single image dehazing, a domain characterized by its ill-posed nature due to the challenge of estimating latent haze-free images from hazy inputs. The DEA-Net employs a specialized attention mechanism combined with an enhanced convolutional architecture to improve feature learning and produce high-quality dehazed images. The authors present a comprehensive methodological framework, including unique architectural components and detailed evaluations.

Core Contributions and Methodology

1. Detail-Enhanced Convolution (DEConv):
   • The paper proposes the DEConv, which integrates difference convolutions with traditional vanilla convolution layers. By embedding traditional prior information through various forms of differential calculations, DEConv enhances the representational and generalization capabilities of the convolution layers. Notably, the DEConv is further optimized by re-parameterizing the layers to avoid extra computational cost, thus maintaining efficiency in deployment.
2. Content-Guided Attention (CGA):
   • The CGA mechanism is a two-step attention generator that provides channel-specific spatial importance maps (SIMs) by leveraging the spatial content of input features. Unlike conventional attention mechanisms that utilize a uniform importance map, CGA generates channel-specific maps in a coarse-to-fine manner, improving the contextual refinement and focus of the network on significant regions across each channel.
3. CGA-Based Mixup Fusion Scheme:
   • A novel mixup fusion approach is introduced to reconcile mismatched receptive fields between low-level and high-level features. By employing CGA to guide the fusion process, this technique effectively adapts the contribution of different feature layers during synthesis, facilitating enhanced feature propagation and gradient flow.

Experimental Evaluation and Results

The DEA-Net demonstrates its superiority through comprehensive experiments on benchmark datasets such as SOTS-indoor, SOTS-outdoor, and Haze4K. Its efficacy is highlighted by achieving a PSNR exceeding 41 dB with only 3.653 million parameters, outperforming several state-of-the-art methods. This performance reflects the robustness of DEA-Net in handling non-uniform haze distributions and extracting salient features necessary for high-fidelity image reconstruction.

Implications and Future Directions

The implications of DEA-Net are significant for both theoretical advancements and practical applications in image processing and computer vision. The introduction of difference convolutions and content-guided attention provides a structured approach to improve feature extraction while maintaining computational efficiency, which is critical for real-time and large-scale applications. The inclusion of re-parameterization techniques points toward an optimization pathway for other neural architectures facing similar trade-offs between model complexity and performance.

Looking forward, DEA-Net establishes a promising precedent for employing well-designed priors and innovative attention mechanisms in convolutional neural networks (CNNs). Future research might explore extending these principles to other complex image restoration tasks or incorporating more dynamic approaches to handle a wider range of environmental conditions, potentially leveraging advances in transformer models to further boost long-range dependency modeling.

In conclusion, the DEA-Net paper offers a substantial contribution to single image dehazing, both through immediate performance gains and by providing a platform for subsequent research extensions in the field.