An Improved DIP Denoising Model Based on Internal and External Image Priors and Image Fusion
To enhance the denoising performance of an unsupervised Deep Image Prior(DIP)model,an improved approach known as the Improved Deep Image Prior(IDIP)is proposed,which comprises sample generation and sample fusion modules,and leverages a prior hybrid image that combines internal and external factors,along with image fusion techniques.In the sample generation module,two representative denoising models are utilized,which capture internal and external priors and process the noisy image to produce two initial denoised images.Subsequently,a spatially random mixer is implemented on these initial denoised images to generate a sufficient number of mixed images.These mixed images,along with the noisy image,form dual-target images with a 50%mixing ratio.Furthermore,executing the standard DIP denoising process multiple times with different random inputs and dual-target images generates a set of diverse sample images with complementary characteristics.In the sample fusion module,to enhance randomness and stability,50%of the sample images are randomly discarded using dropout.Next,an unsupervised fusion network is used,which performs adaptive fusion on the remaining sample images.The resulting fused image exhibits improved image quality compared to the individual sample images and serves as the final denoised output.The experimental results on artificially generated noisy images reveal that the IDIP model is effective,with an improvement of approximately 2 dB in terms of Peak Signal-to-Noise Ratio(PSNR)compared to the original DIP model.Moreover,the IDIP model outperforms other unsupervised denoising models by a significant margin and approaches the performance level of supervised denoising models.When evaluated on real-world noisy images,the IDIP model exhibits superior denoising performance to the compared methods,thus verifying its robustness.
Image denoisingDeep image priorBoosting performanceInternal and external image priorsUnsupervised fusion
万方数据
维普
