In the context of smart cities, edge-aware machine vision systems are widely used. These systems involve scenarios where large volumes of image data are stored locally. They also involve scenarios where image data is uploaded to edge clouds, posing significant privacy risks. Therefore, it is necessary to encrypt images containing sensitive information. However, edge computational devices typically have limited computational ability. To address the need for privacy protection, this paper proposes a partial image encryption algorithm based on object detection. First, our approach uses an object detection model to identify private areas in images (such as license plates) and applies a specific encryption strategy to license plate areas. At the same time, the computational burden on edge devices is reduced. Additionally, we introduce a chaotic mapping algorithm based on image segmentation and compare its performance with traditional chaotic mapping algorithms. Experimental results show that the improved algorithm performs better in encrypting sensitive areas while also exhibiting superior performance in gray value histogram analysis and scatter plot analysis.
The classification of precipitation clouds, particularly the identification of severe convective clouds, is of paramount importance for meteorological forecasting and disaster warning systems. Current precipitation cloud observations typically rely on either standalone satellites or ground-based meteorological stations conducting horizontal-layer detection, yet these methods suffer from limitations such as restricted detection ranges and single-source observation. Therefore, this paper employs multi-source satellite data to construct a vertically structured cloud-precipitation dataset. This dataset comprises four categories: clear skies, non-precipitating clouds, general precipitation clouds, and severe convective clouds. A self-developed DFConv attention mechanism is integrated into the UNet network framework to build the Vertical-UNet model for identifying precipitation cloud types within vertical structures. Experimental results demonstrate that Vertical-UNet achieves favorable performance in precipitation cloud classification using the vertical-structure precipitation cloud dataset. The probability of detection (POD) for precipitation clouds reaches 94.54%. The POD for severe convective clouds reached 87.29%, indicating an improvement of 19.9% compared to the CNN model and 17.04% compared to the UNet model. This conclusively validates the model’s efficacy and establishes a foundation for detecting vertically structured precipitation clouds from diverse satellites in future research.
8 May 2026 in IEEE Transactions on Aerospace and Electronic Systems
Anti-UAV detection has emerged as a critical task in airspace security, yet it remains challenging due to the small size, weak edge responses, and complex backgrounds of UAV targets. For small UAVs occupying only a few pixels, boundary pixels constitute the most informative geometric cues for localization. However, existing methods treat edge extraction as an implicit byproduct of feature learning, causing boundary details to degrade through successive layers without explicit preservation. To address this limitation, we propose EFE-Net, an edge-aware multi-scale feature enhancement network that establishes a unified edge centric pipeline where boundary information is explicitly extracted, selectively enhanced, and systematically propagated throughout the feature hierarchy. Specifically, we design a Frequency-enhanced Multi-Scale Edge Injector (FMSEI) that exploits the complementarity between spatial gradients and frequency-domain spectra to generate robust multi-scale edge representations at the finest pyramid scale. We then propose a Pyramid Contextual Edge Selection (PCES) module that derives selection weights from multi granularity contextual differences through differential operations, adaptively reinforcing boundary activations while suppressing re dundant responses in smooth regions. Furthermore, we introduce a Feature Aggregation Layer (FAL) equipped with the Dual-Domain Cross-Scale Aggregation (DDCSA) module, which applies learnable weights to selectively modulate low-frequency components while preserving high-frequency boundary structures; spatial attention is further incorporated to provide location-specific refinement. Extensive experiments on the DUT-Anti-UAV, Bird-UAV, and Anti UAV infrared (INF) datasets demonstrate that EFE-Net achieves superior performance compared to state-of-the-art methods. Cross domain evaluation on the VisDrone dataset further validates the generalization capability of our approach.
In order to solve the problems of insufficient medical image feature extraction, high classification accuracy, and computational complexity in automatic diagnosis of skin lesions in the edge computing environment, this paper proposes a real-time pseudo-multimodal low-delay diagnosis framework, SCGViT, based on a vision transformer. The framework is constructed around three functional objectives: mitigating data imbalance through generative modeling, capturing diverse representations via multi-dimensional perception, and optimizing feature fusion through adaptive refinement. Firstly, using Class-Conditioned Generative Adversarial Networks (CGANs) simulates manifolds of minority class samples in latent space, achieving preliminary balance of data distribution. Secondly, a branch feature-extraction path is constructed to simulate inversion (INV) and infrared (IR) modes in the original visual primary color mode (RGB), in order to achieve multi-dimensional perception. Finally, a cross-attention mechanism is combined for cross-branch feature aggregation, and a channel-attention mechanism (squeeze and excitation) is embedded for secondary refinement of the mixed global local features to enhance the representation ability of key pathological regions by integrating complementary structural and contrast information. The experimental results on the HAM10000 dataset showed that the F1 score reached 0.973, the inference speed reached 304.439 FPS, the parameter count was only 0.524 M, and the computational complexity was only 0.866 G FLOPs, achieving a balance between high accuracy and light weight.