Abstract: Airborne pulse-Doppler radar plays a crucial role in target detection， environmental perception， and guidance monitoring. However， existing equipment is highly susceptible to multi-source complex electromagnetic interference from wireless communications， broadcasting， and electronic countermeasures within its operating frequency band. Such interference leads to signal distortion， noise amplification， and texture degradation in radar echoes， which severely restricts imaging quality and the accuracy with which motion parameters can be estimated. To address this problem， we propose a two-stage interference suppression and image reconstruction framework that integrates a deterministic model based on a UNet architecture with a denoising diffusion probabilistic model （DDPM）. In the first stage， the UNet model performs preliminary structural restoration and energy redistribution through multiscale feature encoding and skip connections to ensure accurate recovery of the global morphology. In the second stage， a conditional diffusion model introduces a soft prior fusion strategy to enable generative enhancement of local textures and fine details without relying on explicit masks to achieve flexible reconstruction. The proposed method was systematically evaluated under conditions with varying signal-to-noise ratio （SNR） and signal-to-jamming ratio （SJR） using simulated radar datasets containing multiple types of interference including comb-spectrum， noise-convolution interference， spectral diffusion， and slice reconstruction interference. The experimental results demonstrate that the proposed framework significantly outperformed approaches based on a single model in terms of structural consistency， texture fidelity， and perceptual quality. In scenarios with high levels of interference， the method achieved an average improvement in peak SNR （PSNR） of approximately 2.1 dB and a 48% reduction in learned perceptual image patch similarity （LPIPS）. Furthermore， the results of parameter estimation experiments validate that the reconstructed images substantially improved the estimation accuracy of target altitude， velocity， and pitch angle， with the mean absolute percentage error （MAPE） in pitch estimation reduced by more than 45%. These results indicate that the proposed two-stage method effectively balanced deterministic regression and probabilistic generation to achieve an optimal trade-off between structural accuracy and perceptual realism. Thus， the proposed approach provides a robust and efficient solution for radar image restoration and motion parameter inversion under complex interference environments.