This study investigates two denoising strategies for improving decomposition of time reversal microwave imaging under severe noise: signal averaging and a denoising autoencoder. The imaging process is based on singular value decomposition of the multistatic data matrix, where dominant singular values represent the signal subspace associated with physical scatterers. At low signal-to-noise ratios, the signal and noise subspaces merge, which can mask the dominant modes and degrade target detection and image reconstruction. To evaluate the denoising performance, complex additive white Gaussian noise is introduced into simulated multistatic data over a two-dimensional imaging domain containing point-like scatterers. Signal averaging suppresses uncorrelated noise through repeated measurements, whereas the denoising autoencoder learns a mapping from noisy data to the corresponding noise-free representation. Their effects are assessed through singular-value separation and reconstructed decomposition of time reversal images over challenging noise conditions. The results show that both methods improve the recoverability of the dominant signal modes, but the denoising autoencoder provides stronger subspace separation when the noise level is severe. In particular, at a signal-to-noise ratio of minus five decibels, the autoencoder preserves the dominant singular values and enables reliable image reconstruction with substantially fewer repeated observations than signal averaging. These findings indicate that learned denoising can provide an efficient alternative to repeated-measurement averaging for robust time-reversal microwave imaging in low-signal-to-noise environments.