DisRnet: Discrete Relevance semantic segmentation network based on Discrete Data Distribution

Anonymous

Keywords: Discrete Data Analysis, Residual structure, Pool layer, Image features, Related Features, Discrete features.

TL;DR: The network is designed by analyzing the statistical properties of discrete data.

Abstract: In the current pixel-level semantic segmentation tasks, the current popular algorithm models are generally based on CNNs and combine contextual information to achieve semantic segmentation of images. However, when these algorithms extract features, they are all affected by the mean pooling layer or the maximum pooling layer, which causes the extracted features lose some spatial information easily. Aiming at these problems, this paper designs a discrete pooling layer (Dis pool) and a correlation pooling layer (Rel pool) by using the characteristics of discrete data distribution. The Dis pool can retain the spatial location information of features through the discrete characteristics of discrete data. The Rel pool can utilize the correlational information between the discrete data to preserve the correlation between the features. Then, DisRnet is designed by fusing the Dis pool and Rel pool on the residual structure. Finally, under the Cityscapes, SBD datasets and Pascal VOC dasets, compared with some SOTA models, it is verified that DisRnet has superior performance.