DCSR: Dilated Convolutions for Single Image Super-Resolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Dilated convolutions effectively expand receptive fields in deep learning models without increasing parameters or losing resolution.
• Pixel-level prediction tasks, such as image super-resolution (SR), benefit from architectures that capture multi-scale contextual information.

Purpose of the Study:

• To propose a novel multiscale single image super-resolution (SR) method utilizing dilated convolutions.
• To introduce and analyze the effectiveness of 'mixed convolutions' that combine standard and dilated convolutions within a single layer.

Main Methods:

• Employed dilated convolutions to enlarge the receptive field size without additional computational cost.
• Introduced 'mixed convolutions' by concatenating features from standard and dilated convolutions in each layer.
• Theoretically analyzed the receptive field and intensity properties of mixed convolutions for SR tasks.
• Trained and evaluated networks of varying depths (5, 10, and 20 layers) including a deep network for performance comparison.

Main Results:

• Mixed convolutions were shown to effectively remove blind spots and capture correlations between low-resolution (LR) and high-resolution (HR) image pairs.
• The proposed method demonstrated good generalization ability.
• Experimental results indicated that the proposed multiscale SR method outperforms state-of-the-art approaches in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM).
• The method showed particular effectiveness for large scale factors in SR.

Conclusions:

• The proposed mixed convolution approach is effective for multiscale single image super-resolution.
• This method offers improved performance over existing techniques, especially for significant upscaling ratios.
• The ability to jointly learn maps at different scales within a single network contributes to enhanced SR results.