DenseNet Matting Algorithm Based on Embedded Improved SKNet
YanLong Xu, Hui Fan, Jinjiang Li · 2020
This paper proposes a DenseNet model based on embedding improved SKNet. The advantage of our method is that it can't only identify different local image structures, but also effectively enhance the spread of useful features in the network by introducing a channel attention mechanism into the network. In addition, we have augmented our DataSet on the existing public DataSet ;and perform qualitative and quantitative evaluation on our new augmented DataSet. In terms of qualitative comparison, the matting obtained by our network is very consistent with the benchmark matting in processing texture, hair and other details. In terms of quantitative comparison. We use the Sum of Absolute Difference and Mean Squared Error to analyze the results, compared with other methods, there is a big improvement. Experimental results show that the algorithm in this paper is superior to other algorithms in subjective vision and objective quantification, and is more robust.