Line spectrum extraction in spectrograms using a DNN with SAM2 encoding and reverse attention

To address the challenge of extracting weak line spectrum from spectrograms, this paper proposes a line spectrum extraction network (LSE-Net) with a simple structure and excellent performance. This network is built on a classic encoder-decoder architecture. To improve training efficiency with a small number of samples, this paper embeds adapters into each layer of the encoder to achieve parameter-efficient fine-tunin, enabling the model to complete training with only a small amount of data and attain excellent performance. To strengthen the line spectrum feature extraction capability, a multi-receptive field feature fusion module is integrated at the skip connections to capture line spectrum information at different scales; meanwhile, a reverse cross attention module is introduced in deep feature processing to accurately enhance the edge features of line spectra and effectively suppress background noise. Experimental results demonstrate that the proposed LSE-Net outperforms state-of-the-art methods (e.g., Deep Lofargram, Trans-UNet) in low signal-to-noise ratio (SNR) scenarios (ranging from −16 dB to −22 dB). Specifically, at −22 dB SNR, LSE-Net achieves a probability of detection of 0.965, a false alarm rate of 0.373, and a line localization accuracy of 0.786, which are 57.1%, 42.3%, and 85.8% higher than those of Deep Lofargram, respectively. The proposed network provides an efficient and reliable solution for weak line spectrum extraction in spectrograms.