This paper presents a novel technique for 3D browsing of wide‐angle fisheye images using view‐dependent perspective correction (VDPC). First, the fisheye imaging model with interior orientation parameters (IOPs) is established. Thereafter, a VDPC model for wide‐angle fisheye images is proposed that adaptively selects correction planes for different areas of the image format. Finally, the wide‐angle fisheye image is re‐projected to obtain the visual effect of browsing in hemispherical space, using the VDPC model and IOPs of the fisheye camera calibrated using the ideal projection ellipse constraint. The proposed technique is tested on several downloaded internet images with unknown IOPs. Results show that the proposed VDPC model achieves a more uniform perspective correction of fisheye images in different areas, and preserves the detailed information with greater flexibility compared with the traditional perspective projection conversion (PPC) technique. The proposed algorithm generates a corrected image of 512 × 512 pixels resolution at a speed of 58 fps when run on a pure central processing unit (CPU) processor. With an ordinary graphics processing unit (GPU) processor, a corrected image of 1024 × 1024 pixels resolution can be generated at 60 fps. Therefore, smooth 3D visualisation of a fisheye image can be realised on a computer using the proposed algorithm, which may benefit applications such as panorama surveillance, robot navigation, etc.

A distributed fisheye video surveillance system (DFVSS) can monitor a wide area without blind spots, but it is often affected by the viewpoint discontinuity and space inconsistency of multiple videos in the area. This paper proposes a novel real‐time fisheye video mosaic algorithm for wide‐area surveillance. First, by extending the line photogrammetry theory under central projection to spherical projection, a fisheye video geo‐registration model is established and estimated using orthogonal parallel lines on the ground, so that all videos of DFVSS are in the unified reference system to eliminate the space inconsistency between them. Second, by combining the photogrammetry orthorectification technique with thin‐plate spline transformation, a fisheye video rectification model is established to eliminate serious distortion in geo‐registered fisheye videos and align them accurately. Third, the viewport‐dependent video selection strategy and video look‐up table computation technique are adopted to create a high‐resolution panorama from input fisheye videos in real time. A parking lot of about 0.4 km2 monitored by eight fisheye cameras was selected as the test area. The experimental result shows the line re‐projection error in fisheye videos is about 0.5 pixels, and the overall efficiency, including panorama creation and mapping to the ground as texture, is not <30 fps. It indicates that the proposed algorithm can achieve a good balance between the limitation of video transmission bandwidth and the smooth observation requirement of computer equipment for the panorama, which is of great value for the construction and application of DFVSS.

No abstract

The cascade of convolution layers and the end-toend training process facilitate CNN feature extraction and transmission, and promote the success of CNN in image processing. However, the drawback of heavily relying on large-scale highquality training samples restricts its applications. To avoid costly and unrealistic manual annotations for large-scale remote sensing images, existing land cover maps are considered as an alternative to manual annotations, in which noisy labels are inevitable. To alleviate the impact of noisy labels, this paper proposes to improve the consistency feature learning ability of CNNs as a feasible solution in practical land cover mapping. Firstly, an intraclass feature consistency constraint is introduced to maintain the consistency of CNN feature maps for the same class. Then, an inter-iteration feature consistency constraint is employed to guide the network to learn features that are consistent with the whole underlying distribution inside a mini-batch. These two feature consistency constraints work in a cooperative and complementary manner with the traditional cross-entropy, and together improve the consistency feature learning ability of the proposed Feature Consistency Network (FCNet). Experimental results demonstrate the effectiveness of the proposed FCNet. Extensive experiments on different network structures validate the generalization of the proposed feature consistency constraints.