Fast and Accurate Image Super Resolution by Deep CNN with Skip Connection and Network in Network (1707.05425v7)

Abstract: We propose a highly efficient and faster Single Image Super-Resolution (SISR) model with Deep Convolutional neural networks (Deep CNN). Deep CNN have recently shown that they have a significant reconstruction performance on single-image super-resolution. Current trend is using deeper CNN layers to improve performance. However, deep models demand larger computation resources and is not suitable for network edge devices like mobile, tablet and IoT devices. Our model achieves state of the art reconstruction performance with at least 10 times lower calculation cost by Deep CNN with Residual Net, Skip Connection and Network in Network (DCSCN). A combination of Deep CNNs and Skip connection layers is used as a feature extractor for image features on both local and global area. Parallelized 1x1 CNNs, like the one called Network in Network, is also used for image reconstruction. That structure reduces the dimensions of the previous layer's output for faster computation with less information loss, and make it possible to process original images directly. Also we optimize the number of layers and filters of each CNN to significantly reduce the calculation cost. Thus, the proposed algorithm not only achieves the state of the art performance but also achieves faster and efficient computation. Code is available at https://github.com/jiny2001/dcscn-super-resolution

• The paper presents DCSCN, a novel model that reduces layers to 11 using skip connections and 1x1 convolutions for efficient image reconstruction.
• It achieves competitive PSNR and SSIM results on Set5, Set14, and BSDS100 while reducing computational costs by at least 10x compared to traditional models.
• The model’s efficiency makes it ideal for real-time applications on mobile and IoT devices where resource constraints are critical.

Fast and Accurate Image Super Resolution by Deep CNN with Skip Connection and Network in Network

This research paper presents a novel approach to Single Image Super-Resolution (SISR) by utilizing an optimized convolutional neural network (CNN) architecture. The proposed method, termed Deep CNN with Skip Connection and Network in Network (DCSCN), focuses on achieving high-quality image reconstructions with significantly reduced computational costs.

Methodology

DCSCN distinguishes itself from traditional deep learning models, like SRCNN or VDSR, by optimizing network structure to minimize resource consumption without sacrificing output quality. The fundamental innovation lies in the integration of skip connections and the use of a Network in Network (NiN) configuration within the reconstruction network.

Model Structure:

• The DCSCN model comprises 11 CNN layers, compared to the 20-30 layers typical in other state-of-the-art models. This reduction is achieved through the employment of 1x1 convolutions as suggested by the NiN approach, which helps in maintaining a lower-dimensional representation, enabling faster computations.
• Skip connections facilitate the integration of both local and global image features directly into the reconstruction network, enhancing detail retention and overall image quality.

Numerical Results

The DCSCN method is benchmarked against several prominent SISR algorithms, demonstrating competitive performance across the Set5, Set14, and BSDS100 datasets with regards to PSNR and SSIM metrics. Notably, the paper highlights that DCSCN achieves these outcomes with at least a tenfold reduction in computational complexity compared to its counterparts such as VDSR and RED.

Implications

The DCSCN model is particularly relevant in the context of mobile and IoT devices, where computational resources are often limited. Its efficiency makes it practical for real-time applications in environments characterized by constrained bandwidth and processing capabilities.

Potential Future Developments

While DCSCN shows promising results, the authors indicate areas for further research, such as automating the design of model architectures specific to varying SISR tasks. Additionally, exploring ensemble learning models could enhance performance by integrating multiple small-scale deep learning architectures, leveraging their diversity to address complex imaging challenges efficiently.

Conclusion

This paper contributes a methodologically innovative and computationally efficient model for high-performance image super-resolution. The balance between resource demand and reconstruction quality positions DCSCN as a viable solution for modern image processing challenges, particularly on edge devices where efficiency is paramount. As the field progresses, further tuning of model structures and examination of ensemble strategies could yield even more effective solutions for diverse application domains.