A DAC-CLGD-Danet network based method for defaced image segmentation

Abstract

Based on the problems of high noise, lower contrast, and complex features in defaced images and the low accuracy of existing defaced image segmentation techniques, this paper proposes a defaced image segmentation algorithm based on DAC-CLGD-Danet. Firstly, a CBDNet asymmetric blind denoising network is used for noise-containing defaced images, and natural and synthetic images are trained together to model the image noise and enhance the denoising ability of natural noise. Secondly, Danet is used as the base network. A Dense Atrous Convolution module (DAC) is added to the dual attention mechanism module to extend the perceptual domain of deep convolution, reduce image feature loss, and enhance the representation of global information and edge features of defaced images; Cross-Level Gating Decoder module (CLGD) is introduced to lighten the segmentation network, enhance image context aggregation, and produce accurate semantic segmentation. The experimental results demonstrated that the method in this paper has a significant effect on the HRF dataset and Cityscapes dataset, with a significant improvement compared with FCN, UNet, and SETR models, with Intersection over Union (IoU) improved by 9.81% and Mean Intersection over Union (mIoU) improved by 3.01% compared with UNet.

Keywords

• image segmentation
• danet
• deep learning
• neural networks
• defaced images