---
title: Multi-Scale Feature Interaction Network
url: https://www.emergentmind.com/topics/multi-scale-feature-interaction-network-msfin
type: topic
---

# Multi-Scale Feature Interaction Network

to=arxiv_search  大发游戏官网_code
{"query":"id:2103.13028 OR title:\"Lightweight Image Super-Resolution with Multi-scale Feature Interaction Network\"","max_results":5}
to=search_arxiv  мәҗбур_code
{"query":"2103.13028","max_results":5}
The Multi-scale Feature Interaction Network (MSFIN) is a lightweight convolutional network for single-image super-resolution (SISR) that reconstructs high-resolution images from low-resolution inputs while targeting deployment scenarios with limited storage and computing resources. Its defining design combines hierarchical multi-scale feature extraction, explicit inter-scale interaction, and a lightweight recurrent residual channel attention block (RRCAB). In the original formulation, MSFIN is organized into shallow feature extraction, multi-scale feature extraction, and reconstruction, and is trained with an $L_1$ objective [2103.13028].

## 1. Design objective and problem setting

MSFIN was introduced against a background in which deep and complex convolutional neural networks had already achieved promising SISR performance, but often at the cost of higher memory consumption. The network therefore addresses a specific engineering target: lightweight SISR with a favorable balance between reconstruction quality and model compactness, especially for mobile devices and related resource-constrained settings [2103.13028].

The central design rationale is twofold. First, MSFIN expands the receptive field and exploits informative features from the low-resolution observed image at various scales. Second, it couples those scales through interactive connections rather than treating them as isolated branches. The architecture further incorporates channel attention, but does so through a lightweight block design intended to avoid the parameter growth associated with heavier attention-based super-resolution models [2103.13028].

This combination places MSFIN within a line of super-resolution architectures that seek performance gains from representational organization rather than from simply increasing depth or width. In the terminology of the paper, lightweight behavior is achieved through group convolutions, shallow network design, and parameter sharing via recurrence [2103.13028].

## 2. End-to-end architecture and optimization

MSFIN is partitioned into three main parts:

1. **Shallow Feature Extraction**
2. **Multi-scale Feature Extraction**
3. **Reconstruction**

Shallow feature extraction applies a single $3\times3$ convolution to the interpolated low-resolution input to obtain low-level features. The full signal path is expressed as

$$
F_{SF} = H_{sext}(I_{LR}),
$$

$$
F_{DF} = H_{msfe}(F_{SF}) = H_{dext}(H_{msfim}(F_{SF})),
$$

$$
I_{SR} = H_{re}(F_{DF}) = H_{MSFIN}(I_{LR}).
$$

Here, $H_{sext}$ is the shallow feature extractor, $H_{msfim}$ is the Multi-Scale Feature Interaction Module, $H_{dext}$ denotes deep feature extraction after multi-scale interaction, and $H_{re}$ maps the deep representation back into the high-resolution image space [2103.13028].

Training uses an $L_1$ reconstruction loss:

$$
L(\Theta) = \frac{1}{M}\sum_{i=1}^M \|H_{MSFIN}(I_{LR}^j) - I_{HR}^j\|_1.
$$

This formulation is standard for image restoration, but within MSFIN it serves a specifically lightweight architecture whose complexity is concentrated in feature reuse and inter-scale communication rather than in very deep stacks of conventional convolutions [2103.13028].

## 3. Multi-scale Feature Interaction Module

The core of MSFIN is the Multi-Scale Feature Interaction Module (MSFIM), which performs hierarchical feature extraction and feature interaction across three levels:

- **L1**: shallow level
- **L2**: mid-scale level, obtained by $2\times$ downsampling
- **L3**: deepest level, obtained by $4\times$ downsampling

Each level contains several RRCABs. L2 and L3 use downsampling and upsampling operations, with L3 additionally employing progressive upsampling and local feature fusion. In particular, L3 uses intermediate concatenation followed by a $1\times1$ convolution to aggregate features across RRCABs [2103.13028].

The distinctive property of MSFIM is that coarse-scale features are not merely decoded back upward; they are injected into finer scales through explicit interconnections. Outputs from deeper levels are fused into shallower levels by transposed convolution, and element-wise summation plus skip connections reintegrate the resulting representations. The final multi-scale fusion is written as

$$
F_{DF} = H_{DF}(F_{L1} + F_{L2} + F_{L3}),
$$

where $H_{DF}$ denotes a cascade of RRCABs for further feature aggregation [2103.13028].

The paper gives representative interaction equations. At L1,

$$
F_{L1}^2 = H_{L1}^2(H_{L1}^1(F_{SF} + H_{deco}(F_{L2}^1)) + H_{deco}(F_{L2}^2)),
$$

where $H_{deco}$ is a transposed convolution used to inject upsampled L2 features into the L1 stream. At L2,

$$
F_{L2}^2 = H_{L2}^2(H_{L2}^1(F_{d}^2 + H_{deco}(F_{L3}^2)) + H_{deco}(F_{L3}^4)),
$$

where $F_{d}^2$ denotes features obtained by downsampling $F_{SF}$ [2103.13028].

These equations make clear that MSFIN does not restrict information flow to a unidirectional top-down or bottom-up path. Instead, the network organizes feature exchange between coarse and fine resolutions before the final reconstruction layers. A plausible implication is that the model uses coarse levels to enlarge effective context while retaining the finer localization carried by the shallow branch.

## 4. Lightweight Recurrent Residual Channel Attention Block

MSFIN’s basic computation unit is the lightweight recurrent residual channel attention block, or RRCAB. Its purpose is to model channel dependencies and increase nonlinearity without incurring substantial parameter growth [2103.13028].

RRCAB combines four design elements:

- **Group convolution** in place of standard convolution, using 6 groups
- **$1\times1$ convolution** for feature fusion after grouped processing
- **Channel attention** with SE-like reasoning
- **A recurrent structure** that loops the block once to increase depth without extra parameters

The channel attention mechanism is formulated as

$$
F_{ca} = H_{ca}(F_{in}) = F_{in} \times H_{sig}(H_{1\times1}^1(\sigma(H_{1\times1}^2(H_{ga}(F_{in}))))),
$$

where $H_{ga}$ is global average pooling, $H_{sig}$ is sigmoid, and $H_{1\times1}$ denotes $1\times1$ convolution [2103.13028].

The recurrent residual path is

$$
F_m = H_{rec}(F_{m-1} + H_{ca}(H_{1\times1}^3(\sigma(H_{gc1}(\sigma(H_{gc2}(F_{m-1}))))))),
$$

with $H_{gc1}$ and $H_{gc2}$ denoting group convolutions and $H_{rec}$ denoting the recurrence [2103.13028].

Functionally, RRCAB provides a low-overhead way to retain the representational advantages of channel attention in a compact network. The paper’s ablation discussion reports that removing channel attention or feature fusion from RRCAB leads to consistent PSNR reductions, indicating that the block is not merely a parameter-saving device but a core contributor to performance [2103.13028].

## 5. Model variants, benchmarks, and trade-offs

The paper reports two principal model sizes:

- **MSFIN-S**: 352K parameters
- **MSFIN**: 682K parameters

On $\times 4$ super-resolution benchmarks including Set5, Set14, BSDS100, and Urban100, the smaller variant is reported to outperform or match existing lightweight models such as CARN-M and s-LWSR32, while the larger variant is comparable to or slightly better than models such as IMDN, CARN, and s-LWSR64 [2103.13028].

| Model | Params | Set5 PSNR/SSIM ($\times 4$) |
|---|---:|---:|
| CARN-M | 412K | 31.92 / 0.8903 |
| s-LWSR32 | 571K | 32.04 / 0.8930 |
| MSFIN-S | 352K | 32.08 / 0.8934 |
| MSFIN-S+ | 352K | 32.18 / 0.8945 |
| MSFIN | 682K | 32.28 / 0.8957 |

The comparison is not limited to Set5. On Urban100, for example, **MSFIN-S+** reaches **25.97dB / 0.7804**, compared with **25.87dB / 0.7790** for **s-LWSR32**. For the larger model, the paper reports **32.28dB / 0.8957** on Set5 for **MSFIN**, compared with **32.21dB / 0.8948** for **IMDN**. Self-ensemble provides a further increment, with **MSFIN+** reaching **32.39dB / 0.8971** on Set5 [2103.13028].

A particularly concrete trade-off is reported for feature interaction inside MSFIM: introducing feature interaction yields **+0.04dB** PSNR for **+16K** parameters. The ablation study also states that removing channel attention or feature fusion from RRCAB causes consistent PSNR drops. Qualitative comparisons in Fig. 4 are reported to show finer structure recovery and sharper details relative to other lightweight models [2103.13028].

Taken together, these results define MSFIN less as an attempt to maximize absolute reconstruction scores than as an attempt to optimize the accuracy–compactness frontier for lightweight SISR.

## 6. Position within multi-scale interaction research

Within super-resolution, MSFIN is an explicit formulation of the idea that multi-scale representations should interact rather than merely coexist. That view later appeared in other domains under different architectures and terminology. For example, FIPGNet for salient object detection argues that current methods often “simply aggregate different scale features, ignoring the correlation between different scale features,” and proposes attention-based feature interaction strategies to address that limitation [2407.04085]. Similarly, the multistep feature aggregation framework for salient object detection critiques prior “one-way” fusion schemes and introduces multiscale interaction as a separate module [2211.06697].

This suggests that the 2021 MSFIN belongs to a broader research movement in which multi-scale processing is increasingly treated as an interaction problem rather than only as a pyramid-construction problem. In that sense, MSFIN’s use of inter-level connections, recurrent lightweight attention blocks, and final aggregation over $F_{L1}$, $F_{L2}$, and $F_{L3}$ anticipates later cross-scale designs in other vision tasks [2103.13028].

A separate source of confusion is nomenclature. The acronym **MsFIN** was later reused for a traffic accident anticipation model that aggregates short-term, mid-term, and long-term temporal features from dashcam videos [2509.19227]. That later usage refers to a different architecture and task. In the SISR literature, however, **MSFIN** specifically denotes the lightweight image super-resolution network introduced in 2021, built around MSFIM and RRCAB [2103.13028].

Source: https://www.emergentmind.com/topics/multi-scale-feature-interaction-network-msfin