Real-Time Versatile Object Detection With High Accuracy And Reduced Information Loss During Wartime

Real time object detection models can quickly scan camera feeds or drone footage, which automatically detect and classify objects, to identify potential threats like enemy combatants, vehicles, or weapons through constantly updating battlefield pictures. Existing methods ignore the need for open vocabulary detection, and some models struggle with information loss during data processing. In this paper, we attempted to identify real-time, highly versatile, and accurate object detection by its strengths and ability to address information loss by adopting SPPELAN module. The proposal system detects the open vocabulary using CLIP encoder as language vision model and modified C2f-darknet for fusing visual features and vocabulary embeddings and Path Aggregation Network is used to create multiple pathways to focus on different aspects of the screen to achieve higher parameter usage. Proposed a state-of-art system with natural language supervision for vocabulary learning. We evaluated the performance of our system on a publicly available Microsoft Coco dataset. The proposed methodology detects the accurate region according to the given vocabulary without losing data effectively.