In the rapidly evolving field of video surveillance and security, there is an urgent need for facial recognition systems that can autonomously adapt and improve their accuracy in real-time, particularly in dynamic environments where new faces frequently appear, and lighting conditions constantly change. This paper addresses the critical challenge of developing self-learning camera systems capable of continuously enhancing facial recognition performance without manual intervention, a requirement that has become increasingly vital for maintaining effective security measures in various sectors including public safety, retail, and smart cities. To tackle this pressing issue, we propose a novel autolearning method that combines RetinaFace for face detection and ArcFace for face recognition, incorporating a self-updating mechanism. Our approach consists of three key steps: (1) automatic collection of high-confidence facial images (>90% accuracy) during recognition tasks, (2) periodic model retraining using the collected data, and (3) iterative model updating to enhance overall accuracy.
Evaluating the quality of text summarization for Vietnamese is both critical and challenging, particularly as the demand for reliable automatic summarization grows in Vietnamese-language applications. This field has currently been constrained by traditional evaluation methods such as ROUGE and BLEU, which primarily focus on word matching and struggle to assess how effectively a summary conveys meaning and information. Vietnamese presents significant challenges because of its complex morphology, flexible word order, and nuanced semantic relationships. Consequently, traditional metrics often fail to accurately differentiate between summaries of varying quality, complicating robust evaluation. In this study, we apply the FineSurE framework semantics-based evaluation method that leverages large language models to evaluate Vietnamese news summarization. We use the ViMs dataset, comprising 1,945 Vietnamese news articles with expert-annotated summaries and key facts, to benchmark our approach. Experimental results indicate that, although the two summary variants concise and detailed differ significantly in quality, both achieve uniformly high ROUGE scores (all above 92%), which highlights the insensitivity of traditional metrics. In contrast, FineSurE provides much clearer scores, demonstrating that the detailed summaries are significantly more accurate (85.38% compared to 53.12%) and complete (42.39% compared to 13.75%). This underscores FineSurE's superior ability to identify important quality differences in summarizing Vietnamese texts.
Log-based anomaly detection is essential for monitoring large-scale systems, but most deep detectors rely on floating-point computation that is not energy-efficient for edge deployment, and existing spiking alternatives sacrifice either detection accuracy or integer-only execution. This paper presents SpikeLog-v2, a primarily integer-only spiking detector that keeps both. It replaces the recurrent LSTM (Long Short-Term Memory) decoder of the prior spiking pipeline with spike-driven self-attention that captures cross-event dependencies using only integer and logical operations, and extends integer-only execution to the full attention block via bit-shift power normalization. Because this normalization introduces a position-dependent quantization residual, SpikeLog-v2 couples it with an ℓ1-magnitude token-pruning mechanism that drops the positions most affected by the residual, recovering the lost accuracy while further lowering the per-inference spike-operation count. Trained directly in a single stage, without distillation or ANN-to-SNN conversion, SpikeLog-v2 reaches F1 scores of 0.963 (BGL), 0.828 (Spirit), 0.963 (Thunderbird), and 0.834 (HDFS) across four benchmarks and stays stable under anomaly ratios spanning more than an order of magnitude, where batch-statistic normalization collapses. Under 45 nm CMOS operation counting, its projected per-inference energy of 1.944 µJ is 211× lower than the most efficient ANN baseline, while direct measurement on a Raspberry Pi 4 places its per-inference energy on par with the lightest ANN baselines yet at the lowest average power of all methods and 14× below the SpikeLog baseline.