Computing and Informatics

Deep convolutional networks have been widely applied in image denoising tasks with great success. However, many denoising models extract more feature information by increasing the network depth, which does not fully utilize the shallow features, but also makes it difficult to obtain accurate noise information. In this paper, we introduce a novel modified U-Net structure-based boosted dense attention neural network (BDANet) specifically designed for image denoising. The convolutional block within the encoding layer of BDANet incorporates dense connections and residuals, effectively circumventing the vanishing gradient issue through feature reuse and local residual learning. A boosting strategy is employed in the decoding layer to augment residual information in the noise map. To adeptly process edge details in images, BDANet deploys a polarized self-attentive mechanism to direct the densely connected blocks for depth feature extraction. The network is trained with Gaussian noise at random noise levels in the study to make it flexible to handle images with a wide range of noise levels. In experimental comparisons involving additive Gaussian noise, BDANet outperformed conventional denoising networks and attained competitive results relative to state-of-the-art image denoising networks, with an average improvement of approximately 1.03 dB in terms of PSNR values. Visualization results show that the image after denoising by BDANet network is sharper and richer in texture details than other methods.