Research map / Architectures and efficiency / CNN design
Multiscale Low-Frequency Memory Network for Improved Feature Extraction in Convolutional Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes the Multiscale Low-Frequency Memory (MLFM) network, a plug-and-play parallel memory branch of wavelet-based Low-Frequency Memory Units that improves CNN classification on ImageNet and segmentation on Cityscapes.
Problem
CNNs discard low-frequency information as depth increases (favoring high-frequency texture and edge responses), which hurts tasks needing global structure or smooth transitions.
Approach
A parallel low-frequency memory branch of LFMUs, each with input, update, supplement, forget, and output gates, that stores multiscale wavelet low-frequency coefficients and exchanges them with the core CNN at each downsampling node without changing the core architecture.
Key results
- On ImageNet100, MLFM with ResNet18 (L1-L5 placement, Haar wavelet) reaches 81.22% Top-1 vs. 77.86% baseline, a 3.36-point gain.
- On ImageNet-1K, MLFM_ResNet34 improves Top-1 from 73.09% to 74.66%; MLFM_ResNet18 improves from 69.79% to 71.32%.
- On Cityscapes, UNet mIoU rises from 66.25% to 68.32% and FCN_ResNet mIoU from 64.32% to 67.43% with LFMU in encoder and decoder.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MLFM at L1-L5 to ResNet18 | ResNet18 · ImageNet100 | Top-1 accuracy: 77.86% -> 81.22% | Helped |
| Add MLFM to InceptionNeXt | InceptionNeXt · ImageNet100 | Accuracy: 87.58% -> 88.06% | Helped |
| Add MLFM at L1-L5 using Haar wavelets | ResNet18 and ResNet34 · ImageNet1000 | Accuracy: 69.79% -> 71.32%; 73.09% -> 74.66% | Helped |
| Add LFMU to the UNet encoder | UNet · Cityscapes | MIoU: 66.25% -> 67.29% | Helped |
| Add LFMU to the UNet encoder and decoder | UNet · Cityscapes | MIoU: 66.25% -> 68.32% | Helped |
| Add LFMU to the FCN encoder | FCN_ResNet · Cityscapes | MIoU: 64.32% -> 65.27% | Helped |
| Add LFMU to the FCN encoder and decoder | FCN_ResNet · Cityscapes | MIoU: 64.32% -> 67.43% | Helped |
| Add the supplement gate to LFMUs (ablation) | ResNet18 with LFMU placements L1-L2 through L4-L5 · ImageNet100 | Top-1 accuracy: 78.40% -> 78.94%; 78.02% -> 78.66%; 79.88% -> 80.58%; 79.36% -> 81.22%; 78.34% -> 79.36%; 78.52% -> 79.98%; 78.96% -> 80 | Helped |
| Use wavelet decomposition instead of max-pooling or average pooling (ablation) | ResNet18 LFMU forget and supplement gates · ImageNet100 | Top-1 accuracy: 77.64% -> 81.22%; 77.68% -> 81.22% | Helped |
| Vary LFMU placement across network stages (ablation) | ResNet18 · ImageNet100 | Top-1 accuracy: 77.86% baseline; 78.16%–81.22% | Mixed |
Surprising findings
- Haar wavelet, the simplest basis, gives the highest Top-1 accuracy among tested wavelets, while higher-order Daubechies and Coiflet wavelets degrade performance (coif8 drops to 64.74%).
Limitations
- Ablations and main experiments are run on a 100-class ImageNet subset with a single backbone, ResNet18, limiting generality.
- Gains on full ImageNet-1K are smaller and mainly reported for ResNet variants; many CNN results are shown only in figures without tabulated numbers.
- Reported gains are single-run results with no variance or significance analysis.
Experimental setup
Datasets: ImageNet100 ImageNet-1K Cityscapes
Benchmarks: ImageNet100 ImageNet-1K Cityscapes
Metrics: Top-1 accuracy Top-5 accuracy error rate mIoU overall accuracy precision recall
Baselines: ResNet MobileNetV2 SENet RegNet EfficientNet ConvNeXt InceptionNeXt FCN U-Net
Builds on: Feature Pyramid Network wavelet pooling ResNet
Closest papers by meaning
- CNNpack: Packing Convolutional Neural Networks in the Frequency Domain (NeurIPS 2016 · similarity 0.89)
- Masked Frequency Modeling for Self-Supervised Visual Pre-Training (ICLR 2023 · similarity 0.87)
- Learning Across Scales---Multiscale Methods for Convolution Neural Networks (AAAI 2018 · similarity 0.87)
- Learning in the Frequency Domain (CVPR 2020 · similarity 0.86)
- Non-Natural Image Understanding with Advancing Frequency-based Vision Encoders (CVPR 2025 · similarity 0.85)
- Network Memory Footprint Compression Through Jointly Learnable Codebooks and Mappings (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.