Multisource data are captured from different sensors or generated with different generation mechanisms. Ground camera images (images taken from ground-based camera) and rendered images (synthesized by the position information from 3D image-based point cloud) are different-source geospatial data, called cross-domain images. Particularly, in outdoor environments, the registration relationship between the above cross-domain images is available to establish the spatial relationship between 2D and 3D space, which is an indirect solution for virtual-real registration of Augmented Reality (AR). However, the traditional handcrafted feature descriptors cannot match the above cross-domain images because of the low quality of rendered images and the domain gap between cross-domain images. In this paper, inspired by the success achieved by deep learning in computer vision, we first propose an end-to-end network, DIFD-Net, to learn Domain Invariant Feature Descriptors (DIFDs) for cross-domain image patches. The DIFDs are used for cross-domain image patch retrieval to the registration of ground camera and rendered images. Second, we construct a domain-kept consistent loss function, which balances the feature descriptors for narrowing the gap in different domains, to optimize DIFD-Net. Specially, the negative samples are generated from positive during training, and the introduced constraint of intermediate feature maps increases extra supervision information to learn feature descriptors. Finally, experiments show the superiority of DIFDs for the retrieval of cross-domain image patches, which achieves state-of-the-art retrieval performance. Additionally, we use DIFDs to match ground camera images and rendered images, and verify the feasibility of the derived AR virtual-real registration in open outdoor environments.

Establishing the spatial relationship between 2D images captured by real cameras and 3D models of the environment (2D and 3D space) is one way to achieve the virtual–real registration for Augmented Reality (AR) in outdoor environments. In this paper, we propose to match the 2D images captured by real cameras and the rendered images from the 3D image-based point cloud to indirectly establish the spatial relationship between 2D and 3D space. We call these two kinds of images as cross-domain images, because their imaging mechanisms and nature are quite different. However, unlike real camera images, the rendered images from the 3D image-based point cloud are inevitably contaminated with image distortion, blurred resolution, and obstructions, which makes image matching with the handcrafted descriptors or existing feature learning neural networks very challenging. Thus, we first propose a novel end-to-end network, AE-GAN-Net, consisting of two AutoEncoders (AEs) with Generative Adversarial Network (GAN) embedding, to learn invariant feature descriptors for cross-domain image matching. Second, a domain-consistent loss function, which balances image content and consistency of feature descriptors for cross-domain image pairs, is introduced to optimize AE-GAN-Net. AE-GAN-Net effectively captures domain-specific information, which is embedded into the learned feature descriptors, thus making the learned feature descriptors robust against image distortion, variations in viewpoints, spatial resolutions, rotation, and scaling. Experimental results show that AE-GAN-Net achieves state-of-the-art performance for image patch retrieval with the cross-domain image patch dataset, which is built from real camera images and the rendered images from 3D image-based point cloud. Finally, by evaluating virtual–real registration for AR on a campus by using the cross-domain image matching results, we demonstrate the feasibility of applying the proposed virtual–real registration to AR in outdoor environments.

Glaucoma is one of the leading causes of irreversible vision loss globally, often resulting in blindness. Early detection and treatment are critical in mitigating its impact, with retinal fundus imaging being the most common method for early screening. Traditionally, glaucoma is diagnosed by examining structural changes in these images, but this process heavily relies on the subjective judgment of clinicians, which can lead to errors. With the advent of artificial intelligence (AI), computer-aided diagnostic systems have emerged as powerful tools for early glaucoma screening, offering diagnostic accuracy comparable to that of expert ophthalmologists. Accurate glaucoma diagnosis from fundus images hinges on the precise calculation of the optic cup-to-disc ratio, which depends on the accurate segmentation of the optic cup and disc. Given that these regions occupy only a small portion of the fundus image, traditional deep learning methods typically approach segmentation in two stages: first, by locating the optic cup and disc through object detection, and then by performing fine-grained segmentation within the identified regions. However, the effectiveness of these methods is often constrained by the initial detection accuracy. In this paper, we introduce a novel one-stage segmentation framework, GlauSeg-Net, specifically designed for early glaucoma screening using retinal fundus images. Our method leverages a multi-task learning strategy that employs weak labels to pre-locate the segmentation target within feature layers, significantly enhancing the performance of small target segmentation. Experimental results demonstrate that proposed GlauSeg-Net get 96.8% and 88.3% segmentation accuracy on optic disc and optic cup respectively, and outperforms mainstream benchmark methods in segmentation accuracy.

Medical image segmentation plays a crucial role in diagnosis and treatment planning, yet faces persistent challenges including limited annotated data, boundary ambiguity, and high computational demands that hinder clinical deployment. This paper presents DS-AdaptNet, an efficient segmentation framework that addresses these challenges through two synergistic technical innovations. First, we introduce a Multi-Dimensional Morphological Perturbation Augmentation Technique (MD-PAT) that generates anatomically plausible variations through topologically-constrained deformation fields, significantly enhancing training data diversity while preserving critical structural properties. Second, we develop a Context-Aware Adaptive Threshold Optimization (CA-ATO) algorithm that dynamically determines optimal thresholds by integrating multi-scale contextual information and uncertainty estimates, substantially improving boundary delineation accuracy and fine structure preservation. These techniques are integrated with an Efficient Depthwise Convolutional Neural Network (ED-CNN) architecture that employs depth-separable convolutions, dramatically reducing computational complexity while maintaining high segmentation accuracy. Our comprehensive experiments on three benchmark retinal vessel segmentation datasets demonstrate that the proposed DS-AdaptNet achieves state-of-the-art performance while maintaining exceptional efficiency. Notably, our method attains a Dice coefficient of 0.8328 on DRIVE, 0.8110 on CHASE_DB1, and 0.8515 on STARE, consistently outperforming existing approaches. Most importantly, DS-AdaptNet achieves these results with only 1.57M parameters and 44.08 GFLOPs—a 94.9% reduction in parameters and 77.2% reduction in computational operations compared to standard U-Net. These efficiency gains enable real-time retinal vessel analysis on standard hardware without specialized acceleration, making DS-AdaptNet particularly suitable for resource-constrained clinical environments and telemedicine applications. The proposed framework establishes a foundation for developing practical computer-aided diagnostic systems that balance accuracy, efficiency, and clinical utility.