Abstract
In this paper, we develop a concise but efficient network architecture called linear compressing based skip-connecting network (LCSCNet) for image super-resolution. Compared with two representative network architectures with skip connections, ResNet and DenseNet, a linear compressing layer is designed in LCSCNet for skip connection, which connects former feature maps and distinguishes them from newly-explored feature maps. In this way, the proposed LCSCNet enjoys the merits of the distinguish feature treatment of DenseNet and the parameter-economic form of ResNet. Moreover, to better exploit hierarchical information from both low and high levels of various receptive fields in deep models, inspired by gate units in LSTM, we also propose an adaptive element-wise fusion strategy with multi-supervised training. Experimental results in comparison with state-of-the-art algorithms validate the effectiveness of LCSCNet.
Introduction
Single-image super-resolution (SISR) is the task of recovering a high-resolution (HR) image from a single low-resolution (LR) input. This is a classic "ill-posed" problem because multiple high-resolution images can correspond to the same low-resolution one.
Modern deep learning approaches typically rely on two architectures: ResNet (Residual Networks) and DenseNet (Densely Connected Networks).
- ResNet is efficient because it reuses features through simple addition, but it can be restricted in how many "new" features it learns.
- DenseNet explores new features aggressively by concatenating every layer's output, but this leads to a "parameter explosion" as the network gets deeper, making it computationally heavy.
LCSCNet bridges this gap by creating a network that explores features as thoroughly as DenseNet but remains as parameter-efficient as ResNet.
Proposed Architecture
The LCSCNet architecture consists of three main parts: feature extraction, deep feature exploration, and upsampling/reconstruction. The core innovation lies within the deep feature exploration stage, which utilizes the Linear Compressing Unit (LCSCUnit).
1. The LCSCUnit (Linear Compressing Unit)
Instead of choosing between adding features (ResNet) or concatenating them all (DenseNet), the LCSCUnit splits the work into parallel branches. In each unit:
- Linear Compressing (LC) Layer: A $1 \times 1$ convolution "compresses" the incoming feature maps. This acts as an efficient skip connection that preserves essential information from previous layers.
- Nonlinear Mapping: A parallel $3 \times 3$ convolution explores new, complex features.
- Concatenation: The outputs of these two branches are merged.
By adjusting a parameter ($\rho$), researchers can control the ratio of "old" compressed features to "new" explored features, keeping the total channel count constant and preventing the parameter bloat seen in DenseNet.
2. Adaptive Element-wise Fusion
Deep networks produce features at different depths, each with different "receptive fields." Usually, these are fused using simple addition or fixed weights. LCSCNet introduces a Gated Fusion strategy inspired by LSTMs. It learns pixel-wise weights to decide how much each intermediate layer should contribute to the final high-resolution output. This allows the network to adaptively prioritize different features for different parts of an image, such as focusing on edges in one area and textures in another.
Experimental Results
LCSCNet was evaluated on standard SISR benchmarks including Set5, Set14, BSD100, Urban100, and DIV2K. Key findings include:
- Efficiency: In a head-to-head comparison of 34-layer models, Basic LCSCNet (68.9K parameters) outperformed B-DenseNet (219.8K parameters) and ResNet (118.1K parameters) on the Set5 benchmark.
- State-of-the-Art Performance: The "Enhanced" version (E-LCSCNet) achieved competitive PSNR (Peak Signal-to-Noise Ratio) results against heavyweights like EDSR and RDN. For example, at $\times 2$ scale on the Urban100 dataset, E-LCSCNet matched RDN's performance but with significantly fewer parameters (14.2M vs 22.1M).
- Visual Quality: Qualitative results show that LCSCNet recovers sharper edges and more accurate grid structures compared to models like VDSR or MemNet, which often produce blurrier artifacts.
Conclusion
LCSCNet offers a powerful alternative to traditional ResNet and DenseNet architectures for super-resolution. By utilizing linear compression in skip connections and an input-aware adaptive fusion mechanism, it achieves high-quality image reconstruction with significantly reduced parameter overhead. Future work may explore applying the LCSCUnit to other computer vision tasks such as object detection and segmentation.
Summary updated