Underwater images captured in real marine environments usually suffer from wavelength-dependent color attenuation, low contrast, haze-like scattering, and weak structural details. Although supervised underwater image enhancement (UIE) methods have achieved promising results, their performance is still strongly affected by the availability and reliability of paired underwater training data. In addition, commonly used pixel-level or feature-level losses mainly constrain the image modality itself and provide limited guidance from high-level perceptual semantics. To address these problems, this paper proposes PhyCR-UIE, an unsupervised underwater image enhancement framework that integrates physical imaging priors and cross-modal perceptual regularization. Specifically, three CNN-based subnetworks are designed to estimate the global background light, the medium transmission map, and the latent scene radiance in the underwater imaging model. The estimated physical variables are jointly optimized by reconstruction consistency and homology consistency, enabling the network to learn from degraded underwater images without paired clean references. Furthermore, a cross-modal loss based on CLIP-based image quality assessment (CLIP-IQA+) is introduced. By measuring the semantic similarity between the predicted scene radiance and positive and negative image-quality descriptions, this loss guides the enhanced result toward clearer, more natural, and color-balanced visual appearances. Experiments on five real underwater datasets demonstrate favorable color correction, perceptual quality, and underwater image quality ranking.
Virtual reality head-mounted displays (VR-HMDs) impose visual demands that differ from natural viewing conditions, particularly because of the vergence–accommodation conflict generated by immersive stereoscopic displays. This exploratory study evaluated immediate changes in binocular and accommodative visual performance after brief VR-HMD exposure in adults. Using a single-group pre–post design, fifty-four participants underwent optometric assessment before and immediately after 15 min of exposure to a Samsung Gear VR head-mounted display. Near point of convergence, stereoacuity, accommodative facility and near cover test findings were assessed. Pre–post changes were analysed using paired statistical tests and adjusted linear regression models were fitted to explore factors associated with the magnitude of change. After VR-HMD exposure, minor measurable cover test changes were observed in 16 participants, all of 1 prism dioptre and in both exophoric and esophoric directions. Near point of convergence increased slightly, whereas stereoacuity and accommodative facility showed better post-exposure test performance. Baseline stereoacuity and baseline accommodative facility were significantly associated with their respective pre–post changes, while age, spherical equivalent, glasses use and contact lens use were not significant predictors. These findings describe small and heterogeneous changes in visual test performance following brief recreational VR-HMD exposure, without a consistent pattern of deterioration; as no non-VR control condition was included, these changes may reflect VR-related effects, repeated testing or both. From a display-human interaction perspective, objective binocular and accommodative measures may provide useful information when evaluating visual performance, comfort and short-term user response to VR-HMD systems.
3D point cloud compression is crucial for bandwidth-constrained applications such as autonomous driving and augmented reality, where reconstructed geometry supports human-oriented visualization while compressed representations must preserve information for machine perception tasks. However, most existing methods primarily focus on either geometric reconstruction quality or downstream task performance. To address this limitation, this paper proposes a novel framework named Human–Machine Point Cloud Compression via Frequency Decoupling and Adversarial Priors. First, we introduce a Wavelet Spectrum Modulation (WSM) module that processes low-frequency global features and high-frequency local details separately through adaptive gating mechanisms. Second, we leverage adversarially learned structural priors to guide the decoder, encouraging the reconstructed point clouds to better approximate real ones in terms of global structure and distribution consistency. Extensive experiments on ModelNet40 and ShapeNet55 demonstrate that the proposed method achieves Bjøntegaard Delta Rate (BD-Rate) savings of 13.22% and 41.25% over Ulhaq et al. on ShapeNet55 and ModelNet40, respectively, with a maximum saving of 78.58% relative to TMC13 (P=1024) on ModelNet40, while maintaining comparable classification performance.
Car crash detection in road scenarios has long faced challenges including insufficient training samples and limited feature extraction capabilities. From a visual information processing perspective, this problem can be formulated as recognizing traffic incident states from complex visual data. To address the issues above, we propose YOLO-based Car Crash Detection (YOLO-CCD), a detection algorithm based on a dual-teacher knowledge distillation framework that integrates real-world and virtual data sources. The first teacher network leverages a large-scale autonomous driving dataset to enable the student detector to acquire comprehensive representations of normal vehicles and traffic scenes. The second teacher network utilizes virtual accident scenarios to capture accident-specific visual representations and facilitate student detector recognition of accident-involved vehicles. Subsequently, contrastive learning is implemented to analyze feature discrepancies between normal and accident vehicles, enhancing detection reliability. A spatial-channel collaborative attention mechanism is developed to optimize critical feature extraction, adaptively adjusting spatial saliency regions and channel-wise feature importance in the activation maps. To alleviate the data scarcity in real-world accident scenarios, we construct a Virtual Car Crash Detection Dataset (VCCD-12 K) containing diverse accident categories, illumination conditions, and weather variations, which provides comprehensive visual samples for robust detection evaluation. Our framework achieves 54.3% detection accuracy for accident vehicles in complex real-world environments, surpassing current state-of-the-art methods by 4.3 percentage points, while maintaining computational efficiency for practical deployment.
Skin lesion diagnosis relies on complementary information from lesion images and patient-level clinical data. However, in real clinical settings, one modality is often missing or unreliable. Therefore, ensuring robust diagnostic performance under missing-modality conditions is critical for real-world deployment. Existing methods primarily rely on missing-modality reconstruction or knowledge distillation from a full-modality teacher. Many conventional reconstruction-based methods infer the missing representation using only the available modality of the current sample and therefore lack external evidence from similar complete cases. Distillation-only methods transfer knowledge from a full-modality teacher model, but they do not explicitly recover the missing-modality representation. To address these limitations, we propose Retrieval-Augmented Distillation (RA-Distill), a task-specific framework that builds on retrieval-guided recovery, adapts retrieved missing-modality evidence to the current incomplete sample, and aligns the student’s diagnostic predictions with those of a full-modality teacher through logit distillation. The proposed method first uses the available modality to retrieve top-K similar complete cases from a representation bank constructed from the training data. Rather than directly aggregating the retrieved evidence, RA-Distill combines the current incomplete-sample query with each retrieved missing-modality representation before reconstruction. The recovered representation and diagnostic prediction are then optimized through reconstruction supervision and full-modality teacher-logit distillation, respectively. Experiments on three public skin lesion datasets, PAD-UFES-20, Derm7pt, and HAM10000, under both clinical data-missing and image-missing settings show that RA-Distill consistently outperforms representative missing-modality learning methods. These results demonstrate the effectiveness of RA-Distill for robust skin lesion diagnosis under missing-modality conditions. Our code is released on GitHub: https://github.com/Groot-H/RA-Distill.
