Video summarization (VS) aims to generate concise representations of long videos by extracting the most informative frames while maintaining essential content. Existing methods struggle to capture multi-scale dependencies and often rely on suboptimal feature representations, limiting their ability to model complex inter-frame relationships. To address these issues, we propose a multi-scale sensing network that incorporates three key innovations to improve VS. First, we introduce multi-scale dilated convolution blocks with progressively increasing dilation rates to capture temporal context at multiple levels, enabling the network to understand both local transitions and long-range dependencie... More >

Fire detection in smart cities requires intelligent visual recognition systems capable of distinguishing fire from visually similar phenomena while maintaining real-time performance under diverse environmental conditions. Existing deep learning approaches employ attention mechanisms that aggregate spatial information isotropically, failing to capture the inherently directional characteristics of fire and smoke patterns. This paper presents DirFireNet, a novel fire detection framework that exploits directional fire dynamics through Strip Pooling Coordinate Attention (SPCA). Unlike conventional attention mechanisms, DirFireNet explicitly models vertical flame propagation and horizontal smoke d... More >

In response to the challenges of insufficient accuracy in face detection and missed small targets under low-light conditions, this paper proposes a detection scheme that combines image preprocessing and detection model optimization. Firstly, Zero-DCE low-light enhancement is introduced to adaptively restore image details and contrast, providing high-quality inputs for subsequent detection. Secondly, YOLOv11n is enhanced through the following improvements: a P2 small-target detection layer is added while the P5 layer is removed, addressing the original model's deficiency in detecting small targets and streamlining the computational process to balance model complexity and efficiency; the P2 up... More >

As the core component of the copper top-blown smelting, the service life of the injection lance critically affects production stability. To monitor the operating condition of the injection lance, a data-driven model is proposed to predict the Remaining Useful Life (RUL) or service life, namely, the DKT-LSSVM model. Firstly, to reduce noise interference, the Daubechies wavelet with four vanishing moments (DB4) denoising is used to process the raw data. Then, the Kernel Principal Component Analysis (KPCA) method is utilized to extract the principal components from the denoised data, which retains at least 90% information content (18 principal components are obtained). These principal component... More >

Federated learning is a collaborative machine learning paradigm that trains models across multiple computing nodes while aiming to preserve the privacy of local data held by participants. However, because of the open network environment, federated learning faces severe privacy and security challenges. Studies have shown that attackers can reconstruct original training data by intercepting gradients transmitted across the network, thereby posing a serious threat to user privacy. One representative attack is the Deep Leakage from Gradients (DLG), which iteratively recovers training data by optimizing dummy inputs to match the observed gradients. To address this challenge, this paper proposes a... More >