OBJECT DETECTION USING DEEP LEARNING YOLOV8

Abstract

The fast development of deep learning in the last ten years has allowed researchers in the area of computer vision to make great strides in the difficult issue of object identification. Researchers The detection accuracy and inference time of an object detector are usually the deciding factors in its assessment. When comparing detection accuracy, two-stage detectors are often better than single-stage detectors, whereas single-stage detectors have shorter inference times. Nevertheless, there have been notable advancements in the field of object identification and related deep learning tasks, such as the YOLO (You Only Look Once) model, which has greatly improved performance. There are two main types of object detectors: those with two stages and those with one stage. When it comes to possible object detection, two-stage detectors utilise intricate structures that zero down on certain regions, whereas single-stage detectors use more straightforward designs that take into account all geographical regions at once. and other architectural offspring, have greatly enhanced detection accuracy to the point that they may sometimes surpass two-stage detectors. The rapid inference times of YOLO models have led to their widespread use, even though these models often sacrifice detection accuracy for speed. Our suggested system incorporates the cutting-edge YOLOv8 model for identification of weapons in real-time. When compared to YOLOv5, YOLOv8 is much faster and more accurate. We have quantized the YOLOv8 model's weights to guarantee efficient performance. Our trials compared the effectiveness of the YOLOv8 and YOLOv5 versions in detecting weapons. By using YOLOv8, we were able to increase our mean Average Precision (mAP) from 89.1% in YOLOv5 to 90.1%. Additionally, we were able to decrease the inference time by 15% when compared to the original YOLOv8 setup by applying weight quantization to the YOLOv8 model. With its enhanced accuracy and accelerated inference, YOLOv8 is a top pick for applications that need real-time weapon identification.