A Novel Deep Residual Attention Network for Face Mosaic Removal

Abstract

Deep learning used to achieve face mosaic removal is in full swing. In this paper, a novel deep residual attention network (DRAN) is proposed for face mosaic removal. Inspired by the application of attention mechanism, we apply channel attention (CA) and pixel attention (PA) to DRAN to make the network focus on more informative information. In addition, we improve the conventional pixel attention which we superimpose three convolutional kernels of different sizes. DRAN consists of an encoder and a decoder, which the clean and real face image is reconstructed by convolutional neural network. In the encoder, the feature maps of each convolutional layer are used as the input of CA, the output of CA is sent to PA, and the output of PA is directly concatenated with the corresponding feature maps of the decoder. As the same time, inspired by the residual learning, we propose the parallel residual block for more detailed feature extraction. Extensive experiments show that DRAN performs better than state-of-the-art methods, the best PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index) based on the test set are 20.67 dB and 0.8509, respectively.