Multi-Frequency Branch-wise Block (MFB)

• Multi-Frequency Branch-wise Block (MFB) is a design pattern that explicitly separates frequency components into dedicated branches to process high-frequency details and low-frequency structures.
• Architectural variants such as FreConv, T-InvBlock, and MFAF illustrate distinct implementations, using dual or tri-branch mechanisms and different signal processing techniques like Sobel filtering and Haar decomposition.
• Empirical studies demonstrate that employing MFB improves outcomes in image rescaling, visual recognition, and geo-localization by effectively handling frequency-sensitive information.

to=arxiv_search.search 】【。】【”】【json {"query":"(Bao et al., 2024) Plug-and-Play Tri-Branch Invertible Block for Image Rescaling", "max_results": 5} to=arxiv_search.search 】【。】【”】【json {"query":"(Li et al., 2023) FreConv Frequency Branch-and-Integration Convolutional Networks", "max_results": 5} to=arxiv_search.search 】【。】【json {"query":"(Liu et al., 16 Sep 2025) MFAF EVA02-Based Multi-scale Frequency Attention Fusion Method for Cross-View Geo-Localization", "max_results": 5} Multi-Frequency Branch-wise Block (MFB) denotes a class of branch-structured modules that explicitly separate and process signals associated with different frequency regimes before recombination or downstream use. In current arXiv usage, the term is not fully standardized. In "FreConv: Frequency Branch-and-Integration Convolutional Networks," MFB corresponds to a dual-branch convolutional replacement that separates high- and low-frequency information and integrates them by summation. In "MFAF: An EVA02-Based Multi-scale Frequency Attention Fusion Method for Cross-View Geo-Localization," MFB denotes a post-backbone module that produces original, low-frequency, and high-frequency feature branches. In "Plug-and-Play Tri-Branch Invertible Block for Image Rescaling," the paper does not explicitly define a module named MFB, but its T-InvBlock is a concrete multi-frequency branch-wise design because it separates a high-frequency branch from luminance and chrominance low-frequency sub-branches (Li et al., 2023, Liu et al., 16 Sep 2025, Bao et al., 2024).

1. Terminological scope and conceptual basis

The common principle across these formulations is branch-wise specialization over frequency-sensitive content. High-frequency information is treated as the carrier of edges, contours, fine textures, or recoverable detail, while low-frequency information is treated as smoother, more redundant, or more globally stable structure. In FreConv, this motivates a derivative-filter-like high-frequency extractor and a lightweight low-frequency extractor. In MFAF, low-frequency components are described as carrying stable, global layout information such as building footprints and road topology, while high-frequency components carry local details that distinguish visually similar scenes. In T-InvBlocks, the low-frequency component retained for upscaling is further decomposed into luminance and chrominance because processing low-frequency content directly in RGB leaves strong inter-channel correlations unexploited (Li et al., 2023, Liu et al., 16 Sep 2025, Bao et al., 2024).

The term therefore refers less to a single canonical layer than to a design pattern: separate frequency-relevant information into dedicated branches, apply different operators to each branch, and then either integrate the outputs or preserve them as distinct pathways. This suggests that MFB is best understood as an architectural strategy for imposing frequency awareness on otherwise generic convolutional, invertible, or backbone-based pipelines.

2. Architectural variants

Across the cited works, MFB appears in three distinct but related architectural forms (Li et al., 2023, Liu et al., 16 Sep 2025, Bao et al., 2024).

Variant	Branch structure	Distinctive operations
FreConv	Dual branch: low-frequency and high-frequency	Attention-based split, derivative-filter-like HFE, 1×1 LFE, additive fusion
T-InvBlock	Tri-branch: $Y$, $CbCr$, and $HF$	Haar decomposition, RGB→YCbCr on LF, additive LF couplings, affine HF coupling
MFAF MFB	Three outputs: orig, LF, HF	Multi-scale average pooling LF branch, directional Sobel HF branch, FSA refinement

In FreConv, the input feature map $X_{in} \in \mathbb{R}^{C \times H \times W}$ is split by channel-wise attention into a top stream for low-frequency extraction and a bottom stream for high-frequency extraction. The high-frequency branch uses a derivative-filter-like design implemented as a difference between a multi-scale convolution and a point-wise convolution, while the low-frequency branch uses only a point-wise convolution. The two outputs are fused by point-wise summation.

In T-InvBlocks, the branch structure is imposed after Haar wavelet decomposition. The high-resolution image is first decomposed into low-frequency ($x_l$) and high-frequency ($x_h$) parts; the low-frequency part is then converted from RGB to YCbCr and split into luminance ($x_y$) and chrominance ($x_c$). Three branches $[x_y, x_c, x_h]$ are updated jointly by invertible couplings. The luminance and chrominance branches use additive updates, while the high-frequency branch uses affine coupling conditioned on both low-frequency branches.

In MFAF, the block is inserted after the last EVA02-L feature map $F_v \in \mathbb{R}^{H \times W \times C}$. It preserves the original feature map as one branch, constructs a low-frequency branch by multi-scale average pooling with kernel sizes $CbCr$0, $CbCr$1, and $CbCr$2, and constructs a high-frequency branch by directional Sobel filtering followed by channel-attentive fusion. Each branch is then refined by Frequency-aware Spatial Attention (FSA), and the resulting three representations are used independently during training and concatenated during inference.

3. Mathematical structure and signal flow

FreConv formalizes the branch split through global average pooled statistics

$CbCr$3

followed by two gated pathways

$CbCr$4

These gates produce weighted features $CbCr$5 and $CbCr$6, which are split into $CbCr$7 channel groups and summed as

$CbCr$8

The low-frequency extractor is

$CbCr$9

and the high-frequency extractor is

$HF$0

with final integration

$HF$1

The high-frequency branch can enlarge receptive field either by dilated convolution kernels with effective kernel size

$HF$2

or by group convolution kernels chosen so that

$HF$3

This gives FreConv a frequency-aware replacement for a vanilla convolution (Li et al., 2023).

T-InvBlocks extend invertible dual-branch coupling to a tri-branch form. For block $HF$4, forward downscaling is

$HF$5

$HF$6

$HF$7

and inverse upscaling is

$HF$8

$HF$9

$X_{in} \in \mathbb{R}^{C \times H \times W}$0

Here $X_{in} \in \mathbb{R}^{C \times H \times W}$1 are CNN-based functions implemented as densely connected convolutional blocks. Over the full pipeline, the low-frequency luminance and chrominance streams are mapped separately by $X_{in} \in \mathbb{R}^{C \times H \times W}$2 and $X_{in} \in \mathbb{R}^{C \times H \times W}$3, while the high-frequency stream is mapped by $X_{in} \in \mathbb{R}^{C \times H \times W}$4. The inverse stage sets

$X_{in} \in \mathbb{R}^{C \times H \times W}$5

so the model uses deterministic all-zero mapping rather than sampling the high-frequency latent during upscaling (Bao et al., 2024).

MFAF MFB remains entirely in the spatial domain. Its low-frequency branch forms

$X_{in} \in \mathbb{R}^{C \times H \times W}$6

with

$X_{in} \in \mathbb{R}^{C \times H \times W}$7

The high-frequency branch uses four Sobel kernels,

$X_{in} \in \mathbb{R}^{C \times H \times W}$8

to compute per-channel directional responses

$X_{in} \in \mathbb{R}^{C \times H \times W}$9

then concatenates them as

$x_l$0

Channel-wise attention produces

$x_l$1

and the refined high-frequency map is

$x_l$2

FSA is then applied to each branch using Z-pooling and a two-layer $x_l$3 gating network: $x_l$4 The paper explicitly notes that this MFB does not use FFT or DCT (Liu et al., 16 Sep 2025).

4. Application-specific training objectives

The branch structure serves different optimization goals in the three application domains. In T-IRN, the objective is

$x_l$5

with $x_l$6 and $x_l$7. The term $x_l$8 is an $x_l$9 loss in RGB between reconstructed HR and ground-truth HR, and $x_h$0 is an $x_h$1 loss in RGB between downscaled LR and Bicubic LR. Distribution-matching and likelihood terms are removed, and no explicit log-determinant term is used in the loss. T-SAIN retains SAIN’s asymmetric structure, including the compression simulator $x_h$2 and codec $x_h$3, but replaces GMM sampling with all-zero mapping and adds no distribution-related guidance loss (Bao et al., 2024).

In MFAF, the three branches are supervised independently. For each branch $x_h$4, the method uses Cross-Entropy

$x_h$5

and Cross-Domain Triplet Loss

$x_h$6

with margin $x_h$7. The total loss is

$x_h$8

The reported training protocol uses EVA02-L (pretrained), image size $x_h$9, random padding, rotation, cropping, flipping, erasing, SGD with momentum $x_y$0, weight decay $x_y$1, batch size $x_y$2, learning rates $x_y$3 for the backbone and $x_y$4 for other layers, decay by $x_y$5 at epochs $x_y$6 and $x_y$7, and total $x_y$8 epochs on NVIDIA A800 (Liu et al., 16 Sep 2025).

FreConv is trained within standard backbone pipelines rather than with a dedicated frequency-domain loss. Its implementation replaces all vanilla $x_y$9 convolutions in ResNet, VGG, and DenseNet backbones, and replaces max-pooling with strided convolutions to preserve frequency integrity. This reflects a different use of MFB: not as a reconstruction prior or retrieval head, but as an in-place convolutional primitive (Li et al., 2023).

5. Empirical performance and ablation evidence

The empirical record shows that branch-wise frequency specialization is useful in three distinct regimes: image rescaling, generic visual recognition, and cross-view geo-localization (Bao et al., 2024, Li et al., 2023, Liu et al., 16 Sep 2025).

For image rescaling, T-IRN at $x_c$0 improves over IRN with slightly fewer parameters: $x_c$1M versus $x_c$2M. On Set5, PSNR/SSIM improves from $x_c$3 to $x_c$4; on Set14 from $x_c$5 to $x_c$6; on BSD100 from $x_c$7 to $x_c$8; on Urban100 from $x_c$9 to $[x_y, x_c, x_h]$0; and on DIV2K val from $[x_y, x_c, x_h]$1 to $[x_y, x_c, x_h]$2. At $[x_y, x_c, x_h]$3, performance is similar or slightly improved on most datasets. In compression-aware rescaling on DIV2K val, T-SAIN improves over SAIN across all tested JPEG quality factors. At $[x_y, x_c, x_h]$4, for example, QF$[x_y, x_c, x_h]$5 improves from $[x_y, x_c, x_h]$6 to $[x_y, x_c, x_h]$7, and QF$[x_y, x_c, x_h]$8 from $[x_y, x_c, x_h]$9 to $F_v \in \mathbb{R}^{H \times W \times C}$0. At $F_v \in \mathbb{R}^{H \times W \times C}$1, QF$F_v \in \mathbb{R}^{H \times W \times C}$2 improves from $F_v \in \mathbb{R}^{H \times W \times C}$3 to $F_v \in \mathbb{R}^{H \times W \times C}$4, with fewer parameters than SAIN: $F_v \in \mathbb{R}^{H \times W \times C}$5M versus $F_v \in \mathbb{R}^{H \times W \times C}$6M. The ablations are equally important. In T-IRN at $F_v \in \mathbb{R}^{H \times W \times C}$7, simply switching low-frequency processing to YCbCr without tri-branch structure reduces performance relative to IRN, whereas all-zero high-frequency mapping helps, and the full tri-branch plus all-zero mapping performs best. In T-SAIN at $F_v \in \mathbb{R}^{H \times W \times C}$8, low-frequency YCbCr conversion helps under JPEG-like distortions, but the full tri-branch plus all-zero mapping still yields the strongest gains.

For convolutional backbones, FreConv reports simultaneous accuracy and efficiency gains. On ImageNet, Ours-GCK-ResNet-N-2 reduces parameters by $F_v \in \mathbb{R}^{H \times W \times C}$9 ($CbCr$00M vs. $CbCr$01M) and FLOPs by $CbCr$02 ($CbCr$03G vs. $CbCr$04G), with Top-1 improving by $CbCr$05 ($CbCr$06 vs. $CbCr$07). DCK-ResNet50-N-2 reaches Top-1 $CbCr$08 with $CbCr$09M parameters and $CbCr$10G FLOPs. Similar trends are reported for ResNet101, ResNet152, DenseNet121, and VGG16. On COCO with Faster R-CNN, FreConv-ResNet50 improves AP by $CbCr$11 overall and AR by $CbCr$12. On COCO with Mask R-CNN, mask AP improves to $CbCr$13 from $CbCr$14. The ablations attribute these gains to the multi-scale derivative-filter-like high-frequency extractor, the lightweight $CbCr$15 low-frequency extractor, the attention-based split, and point-wise summation as the fusion rule.

For cross-view geo-localization, the gains are large relative to the baseline backbone. On University-1652 with EVA02-L, baseline D2S R@1/AP is $CbCr$16 and S2D R@1/AP is $CbCr$17. Adding MFB raises these to $CbCr$18 and $CbCr$19, and adding FSA yields the full MFAF result of $CbCr$20 and $CbCr$21. On SwinV2-L, the same pattern holds: baseline D2S R@1/AP $CbCr$22 becomes $CbCr$23 with MFB and $CbCr$24 with full MFAF. Branch ablations show that LF + orig outperforms HF + orig on University-1652, but LF + HF + orig performs best overall. The paper also reports top results on SUES-200 and Dense-UAV, and qualitative heatmaps indicating that the low-frequency branch enhances global coarse patterns while the high-frequency branch sharpens edges of buildings, roads, and trees.

6. Relation to adjacent designs, misconceptions, and limitations

A recurrent misconception is that MFB names a single standardized block. The cited literature does not support that interpretation. FreConv uses MFB for a dual-branch convolutional replacement; MFAF uses it for a three-output post-backbone module; and the T-InvBlock paper does not explicitly introduce the term at all, though its tri-branch invertible design fits the same conceptual family (Li et al., 2023, Liu et al., 16 Sep 2025, Bao et al., 2024).

A second misconception is that frequency-aware branch design necessarily requires explicit frequency transforms. MFAF states that its MFB does not use FFT or DCT and instead works in the spatial domain through low-pass averaging and high-pass Sobel filters. FreConv likewise realizes frequency selectivity through architectural constraints rather than explicit spectral transforms. T-InvBlocks use Haar wavelet decomposition and RGB→YCbCr conversion, but their distinguishing contribution is not merely transform choice; it is the tri-branch invertible coupling and deterministic handling of the high-frequency latent.

The ablations also narrow several design claims. In T-IRN, simply converting low-frequency content to YCbCr without introducing a tri-branch architecture does not help under the vanilla invertible block. In FreConv, attention-based integration suppresses performance, and point-wise summation performs best. In MFAF, Z-Pooling with Avg+Max is better than AAP, AP, MP, or Max-only for FSA. These results indicate that branch-wise decomposition alone is insufficient; the interaction rule between branches is a first-order design variable.

The limitations are application-dependent. For FreConv, tasks heavily dominated by low-frequency content may derive smaller gains, and the printed derivative-filter equations are truncated, so the practical implementation rests on the architectural realization of multi-scale minus $CbCr$25 convolution rather than the printed closed form. For T-InvBlocks, exact RGB→YCbCr conversion coefficients are not specified, and FLOPs, memory usage, and runtime are not reported. For MFAF, exact parameter and FLOP counts are also not reported; the low-frequency branch is averaging-based, the high-frequency branch uses fixed Sobel kernels, there is no explicit mid-frequency branch, and cross-branch attention is not learned. The papers themselves point to plausible extensions such as learned low-pass filters, learned high-pass filters, wavelet pyramids, mid-band branches, self-supervised settings, and broader domain transfer. This suggests that MFB remains an evolving design category rather than a closed architectural recipe.