Research on the Extraction of Hazard Sources along High-Speed Railways from High-Resolution Remote Sensing Images Based on TE-ResUNet.

There are many potential hazard sources along high-speed railways that threaten the safety of railway operation. Traditional ground search methods are failing to meet the needs of safe and efficient investigation. In order to accurately and efficiently locate hazard sources along the high-speed railway, this paper proposes a texture-enhanced ResUNet (TE-ResUNet) model for railway hazard sources extraction from high-resolution remote sensing images. According to the characteristics of hazard sources in remote sensing images, TE-ResUNet adopts texture enhancement modules to enhance the texture details of low-level features, and thus improve the extraction accuracy of boundaries and small targets. In addition, a multi-scale Lovász loss function is proposed to deal with the class imbalance problem and force the texture enhancement modules to learn better parameters. The proposed method is compared with the existing methods, namely, FCN8s, PSPNet, DeepLabv3, and AEUNet. The experimental results on the GF-2 railway hazard source dataset show that the TE-ResUNet is superior in terms of overall accuracy, F1-score, and recall. This indicates that the proposed TE-ResUNet can achieve accurate and effective hazard sources extraction, while ensuring high recall for small-area targets.

Regularizing Deep Networks With Semantic Data Augmentation.

Data augmentation is widely known as a simple yet surprisingly effective technique for regularizing deep networks. Conventional data augmentation schemes, e.g., flipping, translation or rotation, are low-level, data-independent and class-agnostic operations, leading to limited diversity for augmented samples. To this end, we propose a novel semantic data augmentation algorithm to complement traditional approaches. The proposed method is inspired by the intriguing property that deep networks are effective in learning linearized features, i.e., certain directions in the deep feature space correspond to meaningful semantic transformations, e.g., changing the background or view angle of an object. Based on this observation, translating training samples along many such directions in the feature space can effectively augment the dataset for more diversity. To implement this idea, we first introduce a sampling based method to obtain semantically meaningful directions efficiently. Then, an upper bound of the expected cross-entropy (CE) loss on the augmented training set is derived by assuming the number of augmented samples goes to infinity, yielding a highly efficient algorithm. In fact, we show that the proposed implicit semantic data augmentation (ISDA) algorithm amounts to minimizing a novel robust CE loss, which adds minimal extra computational cost to a normal training procedure. In addition to supervised learning, ISDA can be applied to semi-supervised learning tasks under the consistency regularization framework, where ISDA amounts to minimizing the upper bound of the expected KL-divergence between the augmented features and the original features. Although being simple, ISDA consistently improves the generalization performance of popular deep models (e.g., ResNets and DenseNets) on a variety of datasets, i.e., CIFAR-10, CIFAR-100, SVHN, ImageNet, and Cityscapes. Code for reproducing our results is available at https://github.com/blackfeather-wang/ISDA-for-Deep-Networks.

High-Resolution Aerial Imagery Semantic Labeling with Dense Pyramid Network.

Semantic segmentation of high-resolution aerial images is of great importance in certain fields, but the increasing spatial resolution brings large intra-class variance and small inter-class differences that can lead to classification ambiguities. Based on high-level contextual features, the deep convolutional neural network (DCNN) is an effective method to deal with semantic segmentation of high-resolution aerial imagery. In this work, a novel dense pyramid network (DPN) is proposed for semantic segmentation. The network starts with group convolutions to deal with multi-sensor data in channel wise to extract feature maps of each channel separately; by doing so, more information from each channel can be preserved. This process is followed by the channel shuffle operation to enhance the representation ability of the network. Then, four densely connected convolutional blocks are utilized to both extract and take full advantage of features. The pyramid pooling module combined with two convolutional layers are set to fuse multi-resolution and multi-sensor features through an effective global scenery prior manner, producing the probability graph for each class. Moreover, the median frequency balanced focal loss is proposed to replace the standard cross entropy loss in the training phase to deal with the class imbalance problem. We evaluate the dense pyramid network on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen and Potsdam 2D semantic labeling dataset, and the results demonstrate that the proposed framework exhibits better performances, compared to the state of the art baseline.