SDWNet: A Straight Dilated Network with Wavelet Transformation for Image Deblurring (2110.05803v1)

Abstract: Image deblurring is a classical computer vision problem that aims to recover a sharp image from a blurred image. To solve this problem, existing methods apply the Encode-Decode architecture to design the complex networks to make a good performance. However, most of these methods use repeated up-sampling and down-sampling structures to expand the receptive field, which results in texture information loss during the sampling process and some of them design the multiple stages that lead to difficulties with convergence. Therefore, our model uses dilated convolution to enable the obtainment of the large receptive field with high spatial resolution. Through making full use of the different receptive fields, our method can achieve better performance. On this basis, we reduce the number of up-sampling and down-sampling and design a simple network structure. Besides, we propose a novel module using the wavelet transform, which effectively helps the network to recover clear high-frequency texture details. Qualitative and quantitative evaluations of real and synthetic datasets show that our deblurring method is comparable to existing algorithms in terms of performance with much lower training requirements. The source code and pre-trained models are available at https://github.com/FlyEgle/SDWNet.

• The paper introduces SDWNet, which combines dilated convolutions to expand receptive fields without extra parameters and wavelet transformation to recover high-frequency details.
• It employs a simplified architecture that accelerates training convergence and reduces computational demands compared to traditional deblurring models.
• Experiments on GoPro, HIDE, and RealBlur benchmarks demonstrate superior PSNR and SSIM performance, validating its high deblurring accuracy.

Overview of "SDWNet: A Straight Dilated Network with Wavelet Transformation for Image Deblurring"

The paper introduces a novel convolutional neural network architecture, referred to as SDWNet, designed to enhance the performance of image deblurring tasks. The authors propose a combination of dilated convolution and wavelet transformation integrated into a streamlined network structure to address common challenges in existing frameworks.

Key Contributions and Methodology

1. Dilated Convolution for Receptive Field Expansion: The authors leverage dilated convolution to achieve a significant improvement in receptive field size without simultaneously increasing the number of network parameters. Unlike traditional Encode-Decode architectures that rely on repeated up-sampling and down-sampling, which can introduce texture detail loss, dilated convolutions maintain spatial resolution and enable efficient capture of non-local features.
2. Wavelet Transformation for High-Frequency Detail Recovery: The integration of a wavelet reconstruction module represents a critical innovation in SDWNet. This module operates in the frequency domain and is utilized in conjunction with the spatial domain to facilitate the recovery of high-frequency texture details, which are often missed by other deblurring techniques. This dual-domain approach ensures preservation of fine details in the reconstructed image.
3. Simplified Network Architecture: SDWNet employs a straightforward architectural design compared to existing deblurring networks. The reduction in complexity aids in faster convergence during training and requires fewer computational resources, making it particularly effective for implementation on systems with limited capabilities.

Experimental Validation and Results

The authors validate the SDWNet on multiple benchmark datasets, including GoPro, HIDE, and RealBlur, demonstrating its competitive performance in both quantitative metrics (PSNR and SSIM) and qualitative assessments. Notably, SDWNet achieves superior performance on the GoPro dataset compared to existing state-of-the-art methods, with significant improvements in SSIM indicating enhanced perceptual quality.

In addition to delivering high deblurring quality, SDWNet is marked by reduced computational demands. The network's parameters and floating-point operations per second (FLOPs) are substantially lower than those of competing models, such as DMPHN and MPRNet, while maintaining or exceeding their deblurring accuracy.

Implications and Future Directions

The introduction of SDWNet suggests several implications for the field of computer vision and image processing. The demonstrated capability of dilated convolution and wavelet transformation to effectively preserve image details while simplifying the network architecture indicates potential applications in related fields such as super-resolution and image denoising. Furthermore, the method's efficiency supports real-time deployment scenarios, critical for video processing or mobile applications.

Looking forward, potential developments might focus on exploring the integration of SDWNet with domain adaptation techniques to improve its generalization on real-world datasets. Additionally, extending the wavelet transformation module to handle varying noise levels in images could further enhance the robustness of the proposed model.

In summary, SDWNet offers a compelling approach for image deblurring tasks, leveraging innovative architectural elements to achieve a balance between performance and efficiency. Its contributions stand to impact both practical applications and theoretical advances in deep learning-based image restoration.