Many image denoising methods achieve state-of-the-art performance with little consideration of computation efficiency. With the popularity of high-definition imaging devices, these denoising methods do not scale well ...Many image denoising methods achieve state-of-the-art performance with little consideration of computation efficiency. With the popularity of high-definition imaging devices, these denoising methods do not scale well to the high-definition images. Therefore, a fast image denoising method is proposed to meet the demand of processing highdefinition images. Based on the analysis of the distribution of the distance distrs between the similar patches and their reference patches from a semantic aspect, the large distrss was found to occur while their contribution to the overall distribution was small. Therefore, the nonlocal filters was replaced in trainable non-local reaction diffusion (TNLRD) with local filters. For image with 4096 × 4096 resolution, the proposed method runs about 6 times faster than TNLRD via a single-thread CPU implementation. And the GPU implementation of the proposed method is about 10 times faster than the CPU implementation. Furthermore, the proposed model achieves competing denoising performance compared with TNLRD in terms of PSNR and SSIM.展开更多
基金the National Key Research and Development Program of China under the Grant No. 2018YFB1003405and National Natural Science Foundation of China under the Grant No. 61732018.
文摘Many image denoising methods achieve state-of-the-art performance with little consideration of computation efficiency. With the popularity of high-definition imaging devices, these denoising methods do not scale well to the high-definition images. Therefore, a fast image denoising method is proposed to meet the demand of processing highdefinition images. Based on the analysis of the distribution of the distance distrs between the similar patches and their reference patches from a semantic aspect, the large distrss was found to occur while their contribution to the overall distribution was small. Therefore, the nonlocal filters was replaced in trainable non-local reaction diffusion (TNLRD) with local filters. For image with 4096 × 4096 resolution, the proposed method runs about 6 times faster than TNLRD via a single-thread CPU implementation. And the GPU implementation of the proposed method is about 10 times faster than the CPU implementation. Furthermore, the proposed model achieves competing denoising performance compared with TNLRD in terms of PSNR and SSIM.
