Semantic segmentation is one of the fundamental tasks in understanding high-resolution aerial images. Recently, convolutional neural network (CNN) and fully convolutional network (FCN) have achieved excellent performance in general images’ semantic segmentation tasks and have been introduced to the field of aerial images. In this paper, we propose a novel deep FCN with channel attention mechanism (CAM-DFCN) for high-resolution aerial images’ semantic segmentation. The CAM-DFCN architecture follows the mode of encoder–decoder. In the encoder, two identical deep residual networks are both divided into multiple levels and acted on spectral images and auxiliary data, respectively. Then, the feature map concatenation is carried out at each level. In the decoder, the channel attention mechanism (CAM) is introduced to automatically weigh the channels of feature maps to perform feature selection. On the one hand, the CAM follows the concatenated feature maps at each level to select more discriminative features for classification. On the other hand, the CAM is used to further weigh the semantic information and spatial location information in the adjacent-level concatenated feature maps for more accurate predictions. We evaluate the proposed CAM-DFCN by using two benchmarks (the Potsdam set and the Vaihingen set) provided by the International Society for Photogrammetry and Remote Sensing. Experimental results show that the proposed method has considerable improvement.

Traffic facilities extraction is of vital importance to various applications such as intelligent transportation systems, infrastructures inventory and city management related applications. Mobile laser scanning (MLS) systems provide a new technique to capture and update traffic facilities information. However, classifying raw MLS point clouds into semantic objects is still one of the most challenging and important issues. In this study, we separate the raw off-ground point clouds into individual segments and explore an object-based Deep Belief Network (DBN) architecture to detect roadside traffic facilities (trees, cars, and traffic poles) with limited labeled samples. To deal with various roadside traffic objects with different types, sizes, orientations and levels of incompleteness, we develop a simple and general multi-view feature descriptor to characterize the global feature of individual objects and extend the quantity of the training samples. Extensive experiments are employed to evaluate the validities of the proposed algorithm with six test datasets acquired by different MLS Systems. Four accuracy evaluation metrics precision, recall, quality and Fscore of trees, cars and traffic poles on the selected MLS datasets achieve (96.08%, 97.61%, 93.86%, 96.81%), (97.55%, 94.10%, 91.69%, 95.58%) and (94.39%, 97.71%, 92.37%, 95.99%), respectively. Accuracy evaluations and comparative studies prove that the proposed method has the ability of achieving the promising performance of roadside traffic facilities detection in complex urban scenes.

Significant progress has been made in the semantic segmentation of mobile laser scanning (MLS) point clouds based on deep learning. However, the segmentation classes of deep learning models depend on the label classes of the source point clouds used for training, which makes it difficult to generalize the models to target point clouds with novel classes. In addition, retraining models using complete class label datasets is time-consuming, and the source point clouds are often unavailable or occupy a large amount of storage space. In this paper, we propose a new class-incremental semantic segmentation framework for MLS point clouds. Firstly, to prevent catastrophic forgetting of original class knowledge when the model learns novel classes, we design a feature representation preservation-based knowledge distillation module to maintain the encoding ability of the target models for original classes. Then, to further separate novel classes from the original background classes, we introduce a background shift mechanism based on loss cross-coupling and pseudo-label collaborative training, which adaptively balances the model plasticity when learning novel class knowledge. Finally, we conducted extensive experiments on two benchmark datasets (Paris-Lille-3D and Toronto-3D), and our proposed method achieved impressive results, which indicate that the proposed framework could effectively achieve class-incremental semantic segmentation for MLS point clouds.