[Paper Review] A heterogeneous group CNN for image super-resolution

Chunwei Tian, Yanning Zhang+4 more|arXiv (Cornell University)|Sep 26, 2022

Image Processing Techniques and Applications4 citations

TL;DR

This paper proposes a heterogeneous group convolutional neural network (HGSRCNN) for single image super-resolution (SISR) that enhances low-frequency structural information through parallel heterogeneous blocks combining symmetric and complementary convolutional units. The method integrates a multi-level enhancement mechanism and a refinement block to suppress redundant features and preserve original details, achieving state-of-the-art performance with competitive inference speed on benchmark datasets.

ABSTRACT

Convolutional neural networks (CNNs) have obtained remarkable performance via deep architectures. However, these CNNs often achieve poor robustness for image super-resolution (SR) under complex scenes. In this paper, we present a heterogeneous group SR CNN (HGSRCNN) via leveraging structure information of different types to obtain a high-quality image. Specifically, each heterogeneous group block (HGB) of HGSRCNN uses a heterogeneous architecture containing a symmetric group convolutional block and a complementary convolutional block in a parallel way to enhance internal and external relations of different channels for facilitating richer low-frequency structure information of different types. To prevent appearance of obtained redundant features, a refinement block with signal enhancements in a serial way is designed to filter useless information. To prevent loss of original information, a multi-level enhancement mechanism guides a CNN to achieve a symmetric architecture for promoting expressive ability of HGSRCNN. Besides, a parallel up-sampling mechanism is developed to train a blind SR model. Extensive experiments illustrate that the proposed HGSRCNN has obtained excellent SR performance in terms of both quantitative and qualitative analysis. Codes can be accessed at https://github.com/hellloxiaotian/HGSRCNN.

Motivation & Objective

• To address the poor robustness of existing deep CNNs in complex scene super-resolution.
• To enhance low-frequency structural information by modeling internal and external channel relations.
• To reduce redundant features through a serial refinement block with signal enhancement.
• To preserve original image details using a symmetric, multi-level enhancement mechanism.
• To enable efficient blind super-resolution training via a parallel up-sampling mechanism.

Proposed method

• HGSRCNN employs a heterogeneous group block (HGB) with two parallel branches: a symmetric group convolutional block and a complementary convolutional block to model diverse channel interactions.
• A refinement block with sequential signal enhancement is applied to filter out redundant features and improve feature quality.
• A multi-level enhancement mechanism enforces a symmetric architecture across stages to progressively strengthen structural representation and expressive capacity.
• A parallel up-sampling mechanism is designed to train a blind super-resolution model without requiring degradation-specific supervision.
• The network is trained end-to-end using a perceptual and adversarial loss to improve visual fidelity.
• The architecture is optimized for inference efficiency, balancing performance and computational cost.

Experimental results

Research questions

• RQ1 Can a heterogeneous CNN architecture improve low-frequency structural representation in super-resolution under complex scenes?
• RQ2 How does combining symmetric and complementary convolutional units in parallel enhance feature learning for SISR?
• RQ3 To what extent does a serial refinement block reduce feature redundancy without losing original image details?
• RQ4 Can a multi-level enhancement mechanism with symmetric design improve feature expressiveness and model robustness?
• RQ5 How does the proposed parallel up-sampling mechanism affect the performance and efficiency of blind super-resolution models?

Key findings

• HGSRCNN achieves a PSNR of 27.60 dB and SSIM of 0.7363 on the DIV80 dataset at ×4 scale, outperforming RDN, CSFM, and SRFBN.
• The model achieves an FSIM score of 0.9785 at ×4 scale on B100, surpassing CFSRCNN and ACNet.
• HGSRCNN runs in 0.0418 seconds for a 1024×1024 image at ×2 scale, representing only 4.27% of SRFBN’s inference time.
• The model’s parameter count (2.178M) and FLOPs (15.05G) are significantly lower than RDN (21.937M parameters, 144.69G FLOPs).
• Visual comparisons show that HGSRCNN produces sharper textures and clearer edges in zoomed regions compared to VDSR, DRCN, CARN-M, LESRCNN, CFSRCNN, and ACNet.
• The multi-level enhancement mechanism enables symmetric feature learning, improving structural consistency and visual quality across scales.