Noise Adaption Network for Morse Code Image Classification

Xiaxia Wang, Xuesong Leng, Guoping Xu

TL;DR

The Noise Adaptation Network (NANet) is a two-stage deep learning framework designed to accurately classify Morse code images even when they are heavily distorted by noise. By decoupling the task into a U-Net-based denoising stage and a multi-scale convolutional classification stage, NANet achieves over 93% accuracy across various noise types, significantly outperforming standard architectures like VGG and MobileNet.

What Problem Does This Paper Address?

While Morse code might seem like a relic of the telegraph era, it remains a vital tool in telecommunications, IoT, and remote positioning due to its simplicity and effectiveness. However, when Morse signals are transmitted or stored as images, they are highly susceptible to environmental noise (Gaussian, uniform, or salt-and-pepper distortions).

Existing classification methods often struggle because they either rely on manual feature extraction—which is slow and brittle—or use end-to-end deep learning models that are typically optimized for a single type of noise. The authors identify a gap in the literature for a robust, automated system capable of handling diverse, multi-source noise interference in Morse code imagery.

Methods

The authors propose NANet, a two-stage architecture that separates "cleaning" the data from "interpreting" it:

  1. Stage 1: Adaptive AutoEncoder: This is a U-shaped (U-Net-like) sub-network. The encoder extracts high-level features while reducing noise through max-pooling, and the decoder reconstructs a "clean" version of the image. A key feature here is the use of skip connections, which help preserve spatial details that might otherwise be lost during the denoising process.
  2. Stage 2: Fine-grained Classifier: The denoised output is fed into a 9-layer deep CNN. To better capture the distinct shapes of Morse "dots" and "dashes," the classifier uses an "expanding convolution kernel" strategy, employing kernels of varying sizes (11x11, 5x5, and 3x3). This allows the network to adapt to different spatial scales of the signal.

Training Strategy: The model uses a joint loss function: TotalLoss = MSELoss (Reconstruction) + CELoss (Classification). This forces the network to learn features that are not only good for denoising but also highly discriminative for identifying the 26 English letters.

Datasets

The researchers introduced a new Multi-noise Morse code image dataset:

  • Core Data: 1,040 pristine images (40 images for each of the 26 English letters) at 512x512 resolution.
  • Noise Variations: To test robustness, they generated three additional versions of the dataset by applying Uniform, Gaussian, and Salt-and-Pepper noise.
  • Availability: The authors have made the datasets available on GitHub.

Key Insights and Ideas

  • Decoupling is Key: Standard models like MobileNetV2 fail spectacularly on uniform noise (dropping to ~7% accuracy). By using a dedicated denoising stage (the Adaptive AutoEncoder), NANet maintains high performance because the classifier always sees a "cleaned" representation.
  • Kernel Diversity: Morse code consists of very short (dots) and relatively long (dashes) features. Using a mix of large (11x11) and small (3x3) kernels in the classifier allows the model to "see" these different signal lengths more effectively than fixed-size kernels.
  • Universal Denoising: In their ablation study, the authors found that placing their Adaptive AutoEncoder in front of other models (like AlexNet or VGG16) improved those models' performance on noisy data, proving the denoising module is a versatile "plug-and-play" component.

Key Results

NANet demonstrated remarkable resilience compared to state-of-the-art (SOTA) models:

Noise Type NANet Accuracy Best SOTA Accuracy (Model)
Uniform 96.92% 87.31% (VGG13)
Gaussian 93.46% 92.69% (VGG11)
Salt-and-Pepper 98.46% 95.00% (VGG11)

Notably, while lightweight models like MobileNetV2 and ShuffleNetV2 collapsed under certain noise conditions (often falling below 10% accuracy), NANet remained stable and highly accurate.

Limitations and Future Work

  • Synthetic vs. Real-World Noise: The study uses mathematically generated noise. Future work could test the model against "real-world" signal degradation, such as motion blur, low lighting, or atmospheric interference in optical Morse transmission.
  • Task Expansion: The authors suggest that this two-stage "denoise-then-classify" approach could be adapted for more complex computer vision tasks, such as object detection and instance segmentation in low-quality imaging environments.
  • Dataset Scale: While effective for the 26 letters, the dataset is relatively small (1,040 images). Scaling this to include numbers, punctuation, or different fonts/styles of Morse representation would be a logical next step.

Summary updated

Abstract

The escalating significance of information security has underscored the per-vasive role of encryption technology in safeguarding communication con-tent. Morse code, a well-established and effective encryption method, has found widespread application in telegraph communication and various do-mains. However, the transmission of Morse code images faces challenges due to diverse noises and distortions, thereby hindering comprehensive clas-sification outcomes. Existing methodologies predominantly concentrate on categorizing Morse code images affected by a single type of noise, neglecting the multitude of scenarios that noise pollution can generate. To overcome this limitation, we propose a novel two-stage approach, termed the Noise Adaptation Network (NANet), for Morse code image classification. Our method involves exclusive training on pristine images while adapting to noisy ones through the extraction of critical information unaffected by noise. In the initial stage, we introduce a U-shaped network structure designed to learn representative features and denoise images. Subsequently, the second stage employs a deep convolutional neural network for classification. By leveraging the denoising module from the first stage, our approach achieves enhanced accuracy and robustness in the subsequent classification phase. We conducted an evaluation of our approach on a diverse dataset, encom-passing Gaussian, salt-and-pepper, and uniform noise variations. The results convincingly demonstrate the superiority of our methodology over existing approaches. The datasets are available on https://github.com/apple1986/MorseCodeImageClassify
BibTeX
@misc{wang2024noise,
  author = {Xiaxia Wang and Xuesong Leng and Guoping Xu},
  title = {{Noise Adaption Network for Morse Code Image Classification}},
  year = {2024},
  eprint = {2410.19180},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.19180},
}
Figure