KJCKorea
Journal Central

In recent large-scale service environments, data pipelines are widely used to collect and transmit web logs in real time. In this process, even when serialization methods such as Avro are applied, repetitive values remain in the data, leading to increased network transmission costs. To address this limitation, this study proposes HAL (Hybrid Analytics Log), a lightweight encoding method that can be integrated into existing systems without significant structural modifications. The proposed approach replaces frequently repeated string combinations with integer identifiers, while preserving unmatched data in its original form to ensure lossless processing. Experimental results using Nginx access logs show that the proposed method reduces Kafka message size by up to 16.2%, Kafka payload bandwidth by 17%, and data ingestion time by 24.6%, without any significant degradation in total processing time or throughput. In addition, the Parquet storage size is further reduced by 4.8% in the final storage stage. These results demonstrate that the proposed approach effectively improves both transmission and storage efficiency through low-cost computational operations.

In this paper, we proposes data preprocessing and augmentation techniques to enhance object detection performance in traffic CCTV systems under various weather and illumination conditions. To overcome the limitations of existing models, which exhibit degraded recognition rates in low-light and adverse weather conditions, we introduce an optimization method for YOLOv8-based models by integrating class balancing, CLAHE(Contrast Limited Adaptive Histogram Equalization), Gaussian blur, and specialized environmental simulations. The proposed method resolves class imbalance through quality score-based downsampling and ensures dataset diversity by simulating various scenarios such as rain, fog, and nighttime conditions. Experimental results demonstrate that the proposed model achieved an F1-score of 0.8792 and an mAP@0.5 of 0.5549, representing a performance improvement of approximately 30.6% and 87.9%, respectively, compared to the baseline model. These results indicate a significant enhancement in the system stability and generalization capabilities required for real-time traffic monitoring environments. Consequently, this research is expected to provide positive contributions to the advancement of smart city infrastructures and traffic safety systems.

In Domain-Incremental Learning (DIL), models must adapt to domain shifts while mitigating catastrophic forgetting. While Pre-Trained Models (PTMs) are widely used, existing strategies often bias representations toward seen domains, failing to effectively mitigate domain shift in unseen domains and inducing representation drift. To address this, we propose a dual-encoder framework that learns domain-invariant and class-discriminative representations. A domain-specific encoder extracts prototypes, distilling stable knowledge into a trainable domain-invariant encoder. Prototype alignment and weight regularization prevent drift and preserve past knowledge. Furthermore, Representation squeezing and a Gradient Reversal Layer (GRL) form compact intra-class clusters and explicitly encourage domain-invariant learning. Extensive experiments on four benchmark datasets (Office-Home, DomainNet, CORe50, and PACS) show our framework improves unseen domain average accuracy by 1-6% across datasets. Notably, on Office-Home, it achieves the highest overall performance with 86.32% Last and 86.25% Avg accuracy. It also reaches an 83.17% Unseen Avg and successfully reduces the forgetting measure to 3.83%.