I/Q Denoising

A transform-domain technique that decomposes the noisy IQ stream into wavelet coefficients, suppresses coefficients below a calculated threshold, and reconstructs the signal. Unlike linear filters, it preserves sharp transient edges and phase discontinuities common in digital modulations.

• Hard Thresholding: Sets coefficients below the threshold to zero, preserving amplitude.
• Soft Thresholding: Shrinks all coefficients toward zero, resulting in a smoother reconstruction.
• VisuShrink: Uses a universal threshold proportional to the noise standard deviation.

A frequency-domain method that estimates the noise power spectrum during signal-free periods and subtracts it from the magnitude spectrum of the noisy IQ segment. The cleaned magnitude is recombined with the original noisy phase to reconstruct the time-domain signal.

• Oversubtraction Factor: Multiplies the noise estimate to reduce residual 'musical noise' artifacts.
• Spectral Floor: A minimum magnitude value to prevent negative spectral components.
• Highly effective for stationary Additive White Gaussian Noise (AWGN).

An optimal linear time-invariant filter that minimizes the Mean Squared Error (MSE) between the estimated clean signal and the true transmitted IQ samples. It requires knowledge of the signal and noise power spectral densities to compute the frequency-domain transfer function.

• Adapts its response based on the local Signal-to-Noise Ratio (SNR).
• Attenuates frequencies where noise dominates; passes frequencies where the signal is strong.
• Often implemented iteratively for non-stationary noise environments.

A subspace-based method that projects a matrix of time-lagged IQ vectors onto its principal components. The signal energy is assumed to be concentrated in the first few eigenvectors associated with the largest eigenvalues, while noise is spread across all dimensions.

• Reconstruction uses only the dominant principal components, discarding the noise subspace.
• Particularly effective for narrowband signals in white noise.
• Does not require explicit frequency-domain transformation.

A data-driven approach using a neural network trained to reconstruct clean IQ samples from corrupted inputs. The encoder compresses the noisy signal into a latent representation, and the decoder reconstructs the denoised output.

• Convolutional Neural Networks (CNNs) capture temporal structure in the IQ stream.
• Denoising Convolutional Neural Networks (DnCNNs) learn residual noise mappings.
• Trained on paired datasets of clean and noisy synthetic IQ, they can model complex, non-linear noise distributions.

A non-linear denoising technique that decomposes the I/Q stream into time-frequency wavelet coefficients. Coefficients falling below a calculated threshold are assumed to be noise and are zeroed out, while significant coefficients representing the signal structure are retained. This method excels at preserving transient edges and impulsive features that linear filters often smear, making it highly effective for pulsed and burst-mode signals.

The fundamental noise model that I/Q denoising algorithms are designed to combat. AWGN adds a random signal with a flat spectral density and Gaussian amplitude distribution to the complex baseband samples. It models the thermal noise generated by receiver electronics. The effectiveness of a denoising algorithm is typically measured by the Signal-to-Noise Ratio (SNR) improvement achieved in the presence of AWGN.

A linear technique that maximizes the SNR of a known signal in the presence of additive stochastic noise. By correlating the received I/Q stream with a time-reversed replica of the expected pulse shape, it acts as an optimal denoising step prior to symbol detection. While not a blind method, it is the theoretical gold standard for pulse-shaped communication signals.

A subspace-based method that separates signal from noise by exploiting energy concentration. The I/Q data matrix is decomposed, and only the principal components corresponding to the highest eigenvalues—assumed to represent the signal subspace—are retained. This effectively discards the noise subspace, making it particularly useful for suppressing white noise in signals with strong temporal correlation.

A technique that uses a reference noise source correlated with the interference in the primary I/Q stream. An adaptive filter dynamically adjusts its coefficients to minimize the mean square error, subtracting the estimated noise from the corrupted signal. This is highly effective for removing narrowband interference or power supply hum that traditional fixed filters cannot track.

A digital smoothing filter that fits successive low-degree polynomials to local frames of I/Q data via linear least squares. Unlike simple moving averages, it excels at preserving the higher-order moments of the signal, such as the shape and width of spectral peaks, while reducing high-frequency noise. It is particularly useful for smoothing instantaneous amplitude and frequency estimates derived from noisy I/Q samples.