Radar image restoration is often challenged by system blur and speckle noise, which degrade target visibility and structural fidelity. In this work, we evaluate three restoration approaches—Richardson–Lucy deconvolution with total variation regularization (RL+TV), Deep Image Prior (DIP), and a new attention-enhanced framework combining DIP with the Convolutional Block Attention Module (DIP+CBAM). A point-spread function (PSF) was extracted from a measured corner reflector and used to simulate realistic degraded radar images. Experimental results show that RL+TV provides limited recovery of fine details, achieving a PSNR of 14.66 dB and SSIM of 0.4969. DIP substantially improves reconstruction quality (PSNR 19.39 dB, SSIM 0.5267), benefiting from the implicit prior of untrained networks. The proposed DIP+CBAM method further enhances performance, reaching the highest PSNR (19.54 dB) and SSIM (0.5361). These findings demonstrate that integrating attention mechanisms into DIP offers a more effective prior for radar image restoration and leads to clearer, more structurally accurate reconstructions.