1results about How to "Improve denoising performance" patented technology

This invention belongs to the field of image processing technology and discloses an image denoising method based on a dual-output regularized network using a composite Jacobi loss. The method includes: first, performing data augmentation, resizing, and normalization on the training dataset and the image to be denoised to generate a noisy image; then, constructing a dual-output FFDNet regularized network, which integrates four functional modules: noise awareness (NA), multi-scale feature fusion (MSFF), hybrid attention (CBAM), and global-local feature enhancement (GFMM-LFEM). This network simultaneously outputs intermediate feature maps and reconstructed images; next, designing a composite loss function including an MSE fidelity term, an L1 regularization term, and a Jacobi regularization term, constraining the monotonicity of pixel changes through the Jacobi matrix; subsequently, training the dual-output FFDNet regularized network using the Adam optimizer and saving the optimal model on the validation set; finally, embedding the trained network as a nearest neighbor denoising module into the PnP-ISTA iterative framework, and iteratively optimizing to output the final denoised image.