Research map / Label-efficient and robust learning / Recognition and long-tail
IncepSeqNet: Advancing Signal Classification with Multi-Shape Augmentation (Student Abstract)
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 IncepSeqNet, a CNN with Inception modules and a multi-shape augmentation (imagification into several 2D shapes) for classifying Zadoff-Chu DMRS sequence signals in real 5G data across SNR levels.
Problem
Classifying sequence signal data (complex-valued Zadoff-Chu DMRS sequences) for 5G systems, where image-oriented CNNs and conventional image augmentations perform poorly, especially at low SNR.
Approach
Combines an Inception module with a Multi-Shape Augmentation step that reshapes the imagified 144-element complex sequence into five input shapes (square 2D format plus (1,S,2)-style shapes), so convolutions with different filter sizes extract more features from the signal.
Key results
- IncepSeqNet achieved better classification accuracy than ResNet18, GoogleNet and InceptionV4 at all eight SNR levels.
- At the lowest SNR (-4.11 dB) IncepSeqNet improved accuracy by about 5 percentage points over InceptionV4.
- Average training time per SNR was 228.9 s for IncepSeqNet versus 305.5 s (InceptionV4), 632.8 s (GoogleNet) and 886.2 s (ResNet18).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Multi-Shape Augmentation | IncepSeqNet · DMRS index classification at -4.11dB | accuracy: 5%p | Helped |
| Use IncepSeqNet instead of existing CNN models | DMRS index classification across eight SNR levels · DMRS | accuracy | Helped |
| Use IncepSeqNet instead of existing CNN models | DMRS training · DMRS (average training time per SNR) | training time: approximately 1.3 to 3.87 times | Helped |
Surprising findings
- Conventional image augmentation techniques degraded performance when applied to imagified sequence signal data.
Limitations
- Evaluation limited to a single proprietary-style DMRS dataset with eight indices and no public release.
- Brief student abstract with limited methodological detail on the augmentation shapes and training setup.
- Only CNN baselines compared; no comparison to other signal classification approaches.
Experimental setup
Datasets: DMRS sequence signals from real 5G base stations (Han, Jo, and Kim 2022)
Benchmarks: DMRS index classification accuracy across eight SNR levels
Metrics: classification accuracy average training time per SNR (s)
Baselines: ResNet18 GoogleNet InceptionV4
Builds on: Inception module GoogLeNet ResNet
Closest papers by meaning
- Robust Automatic Modulation Classification with Fuzzy Regularization (ICML 2025 · similarity 0.78)
- Abstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series Classification (NeurIPS 2024 · similarity 0.77)
- M3Net: Efficient Time-Frequency Integration Network with Mirror Attention for Audio Classification on Edge (AAAI 2025 · similarity 0.77)
- CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching (ICLR 2025 · similarity 0.77)
- Audio Scanning Network: Bridging Time and Frequency Domains for Audio Classification (AAAI 2024 · similarity 0.76)
- Learning Soft Sparse Shapes for Efficient Time-Series Classification (ICML 2025 · similarity 0.76)
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.