A Feature Enhancement-Based Network Framework for Single-Channel Overlapping Signal Source Count Estimation and Time-Frequency Localization

Wireless communication receivers encounter significant challenges in multi-signal environments, particularly with time-frequency overlaps from mixing non-targeted and target signals. This paper designs a new feature enhancement-based network framework for single-channel time-frequency overlapping signal (TFOS) source count estimation combined with time-frequency localization, thus providing critical parameter information for subsequent interference suppression. We employ extended-empirical mode decomposition (Ex-EMD) to construct a single channel into a virtual multi-channel, which decomposes the signal into multiple components and enriches the feature set. Then, we design a multi-scale attention combined global response normalization (MSACGRN) network to implement the estimation of source numbers. MSACGRN network integrates features at different scales and enhances the representation of time-frequency overlapping features. The above process provides information on the number of mixed signals. Next, we construct a time-frequency localization algorithm to distinguish the received overlap from the target of interest and detect the signal positions to supplement the missing time-frequency information. We characterize the time-frequency properties of the signal using the time-frequency image processed by a dual-core filter, which smooths out the noise and enhances the signal edges. Then, we construct the time-frequency localization algorithm by using a channel shuffling network (YOLOCS) based on you only look once (YOLO), where the shuffling module facilitates the exchange of information between the channels and adequately learns the characteristics of the time-frequency interleaving. Results demonstrate that the proposed scheme outperforms existing methods in accurately estimating source counts and time-frequency localization in various overlapping scenarios.