Abstract:
"ABSTRACT
Identifying and recognizing objects in uncertain environments is one of the major challenges in computer vision domain which has been a research challenge for the past decade of time. The performance of existing models and techniques are computationally intensive while it requires the high-end performance. Most of the recognition techniques struggling on degraded and corrupted footages such as light imbalance, data noise, low resolution, blueness and footages in raining conditions. As the recognition systems and computer vision domain is utilizing in various advanced fields such as, law enforcement, autonomous vehicles, robotics etc. Due to that reason there’s a crucial requirement to develop a devise and resilient recognition technique.
The challenges in existing recognition techniques and surveillance systems were firstly identified through a thorough research. In order to overcome the above-mentioned challenges, the proposed VigilanceNet is a novel approach towards optimized recognition model will automatically apply the specific enhancement techniques in order to improve the footage based on the consisting conditions. The YOLOv8 based proposed architecture is optimized as a light weight detection model in order to achieve the computational intensity. While it optimized with enhanced facial and number plate recognition techniques by passing the footage through the stack of down-sampling and up-sampling layers, adjusting their features and hyperparameters as part of the refinement process for detection.
During the final implementation phase, the author conducted the training process on YOLOv8 with the Global Attention Mechanism for number plate detection and recognition by utilizing two datasets for character and number recognition. After conducting multiple training sessions, the author found the optimal number of 100 epochs for number plate dataset and 120 epochs for character recognition dataset. After the successful training process the VigilanceNet model archived a mAP50 & mAP 50-95 value in the range of above 81% to 99% on the Character Recognition Dataset. And for License Plate Image Dataset the a mAP50 & mAP 50-95 value in the range of above 80% to 97%. These results are providing essential information about the performance of the VigilanceNet system on related environments."
