Initial Release

Spectrogram Segmentation Notebook v0.1.0 - Release Notes

Our initial release of the Spectrogram Segmentation Notebook. 🎉

In this example, we use PyTorch and Lightning to train a deep learning model to identify and differentiate between 5G NR and 4G LTE signals within wideband spectrograms.

The machine learning model used in this example is a DeepLabV3 model with a MobileNetV3 large backbone. The dataset used is the Spectrum Sensing dataset provided by MathWorks.

🌟 Notebook Contents

• Background: Delve into the problem background and learn more about the machine learning frameworks, tools, and datasets used in this example.

• Set-up: Install the libraries necessary to run the code in this notebook.

• Data Preprocessing: Load and analyze the Spectrum Sensing dataset.

• Model Training: Configure and train a Deeplabv3 model with a MobileNetV3 backbone.

• Model Validation: Assess the performance of the model using a suite of common machine learning metrics.\n",

• Challenge Data: Challenge the model on combined frames containing both LTE and NR signal.

• Conclusions & Next Steps: Interpret the results, summarize key learnings, and identify steps for expanding upon this example.

🚀 Results

Note: All results are based on the default training and test configurations.

Results on validation dataset

+---------+------------+----------+-------------+------------+-------+
|  Class  |  Accuracy  |  Recall  |  Precision  |  F1 Score  |  IoU  |
+=========+============+==========+=============+============+=======+
|  Noise  |    0.95    |   0.95   |    0.98     |    0.97    | 0.93  |
+---------+------------+----------+-------------+------------+-------+
|   NR    |    0.92    |   0.92   |    0.89     |    0.90    | 0.82  |
+---------+------------+----------+-------------+------------+-------+
|   LTE   |    0.97    |   0.97   |    0.87     |    0.92    | 0.85  |
+---------+------------+----------+-------------+------------+-------+
|  Mean   |    0.95    |   0.95   |    0.91     |    0.93    | 0.87  |
+---------+------------+----------+-------------+------------+-------+

Results on combined (challenge) frames:

+---------+------------+----------+-------------+------------+-------+
|  Class  |  Accuracy  |  Recall  |  Precision  |  F1 Score  |  IoU  |
+=========+============+==========+=============+============+=======+
|  Noise  |    0.92    |   0.92   |    0.71     |    0.80    | 0.67  |
+---------+------------+----------+-------------+------------+-------+
|   NR    |    0.79    |   0.79   |    0.90     |    0.84    | 0.73  |
+---------+------------+----------+-------------+------------+-------+
|   LTE   |    0.58    |   0.58   |    0.91     |    0.70    | 0.54  |
+---------+------------+----------+-------------+------------+-------+
|  Mean   |    0.77    |   0.77   |    0.84     |    0.78    | 0.65  |
+---------+------------+----------+-------------+------------+-------+

🚧 Known Issues

There are no known issues at the time of release.

💡 Future Plans

• We intend to upload this example to Google Colab to increase accessibility, so people don't have to clone the project and download the dataset to get it running.

ℹ️ Additional Information

• We're excited to share this resource under the MIT License, helping others to accelerate your own intelligent radio development and research.
• When evaluating model capacity, it is important to keep in mind that it is an efficient and compact MobileNetV3 model, which has been demonstrated to perform machine vision tasks on lightweight processors such as the Raspberry Pi.