DeepLofargram: A deep learning based fluctuating dim frequency line detection and recovery.

Area of Science:

• Signal Processing
• Machine Learning
• Astronomy

Background:

• Detecting and recovering dim frequency lines in lofargrams is challenging, especially for irregularly fluctuating signals.
• Traditional methods like time integration have limitations in enhancing detection for such signals.
• Deep learning excels at complex visual inference, learning high-level representations from data.

Purpose of the Study:

• To develop an advanced method for detecting and recovering dim frequency lines in lofargrams.
• To leverage deep learning to overcome the limitations of traditional signal processing techniques.
• To achieve state-of-the-art performance in low signal-to-noise ratio (SNR) environments.

Main Methods:

• Proposed DeepLofargram, integrating a deep convolutional neural network with a visualization component.
• Employed a specifically designed multi-task loss function for joint end-to-end training.
• The network learns to detect and recover the spatial location of faint frequency lines.

Main Results:

• Achieved performance limits at signal-to-noise ratios (SNR) as low as -24 dB on average and -26 dB in some cases.
• Demonstrated detection capabilities far exceeding human visual perception.
• Significantly improved upon the current state-of-the-art in lofargram analysis.

Conclusions:

• DeepLofargram effectively detects and recovers dim frequency lines in lofargrams, even at extremely low SNRs.
• The deep learning approach offers a substantial advancement over conventional methods.
• This technique has the potential to revolutionize the analysis of lofargram data.