Semantic Image Segmentation Based on Superpixel-Pooling DCNN With Random Walks on Graph and Application to Intrusion Detection of Railway

Abstract

Deep convolutional neural networks(DCNN) has recently significantly pushed the state of the art of semantic image segmentation. However, as the network depth increases, the progressively lower resolution leads to a loss of feature information and a large number of segmentation errors near the image edges. This is an imperative problem for applications that require accurate segmentation such as intrusion detection of railway. To address this problem, we propose superpixel pooling DCNN combined with random walks networks (SDR-Net), a semantic segmentation model that combines improved DeepLabV3+ and random walks algorithm. Firstly, the improved DeepLabV3+ with superpixel-pooling ASPP (SPASPP) is used as the basic structure of SDR-Net to introduce image prior information and improve the image feature extraction performance. Then, all pixels of the initial segmentation are classified by a binary classification network based on D2S transformation into two categories: correctly labeled pixels and incorrectly labeled pixels. Thirdly, the correctly labeled pixels are used as seed points, and then random walk algorithm is applied to reclassify the mislabeled pixels to achieve higher precision semantic image segmentation. Experiments on PASCAL VOC2012 datasets show the effectiveness of the algorithm in this paper, and it was successfully applied to rail intrusion detection.

Keywords

• Semantic segmentation
• deep convolutional neural networks
• random walks
• superpixel-pooling