DisCnet: Pixel segmentation based on discrete features

Abstract

In the process of pixel-level semantic segmentation tasks, traditional image processing algorithms will suffer from the working mechanism of convolutional layers and pooling layers, resulting in the loss of some features, which now leads to inaccurate semantic segmentation accuracy. For such problems, we design a discrete pooling layer by analyzing the distribution and statistical properties of discrete data. Compared with the traditional pooling layer, the discrete pooling layer can not only preserve the spatial information of features, avoid the loss of features, but also can efficiently improve the accurate segmentation of instance images. Then, based on the discrete pooling layer, we design DisCnet in combination with the convolutional layer. Finally, DisCnet is compared with some state-of-the-art algorithms under the Cityscapes dataset. Experiments demonstrate that DisCnet achieves excellent results in both accuracy and speed.