HIGH-PRECISION OBJECT DETECTION METHOD BASED ON YOLO WITH DYANAMIC THRESHOLDING

Abstract

Object detection systems using the YOLO (You Only Look Once) architecture, through the incorporation of a dynamic thresholding mechanism. The method addresses the challenges of varying object sizes, lighting conditions, and scene complexities by dynamically adjusting detection thresholds in real-time. This adjustment is based on key image characteristics such as brightness, contrast, and object density, which are analyzed before the detection phase. The system also includes a contextual analysis module that refines threshold settings by considering factors like object proximity, typical object sizes, and background complexity, enhancing the model's ability to distinguish between objects and noise. The invention provides significant improvements in detection accuracy by reducing both false positives and false negatives, which are common issues in traditional fixed- threshold object detection systems. This method retains the speed and efficiency of the original YOLO architecture, making it suitable for real-time applications. Moreover, the dynamic thresholding mechanism is scalable and can be applied to various YOLO-based models and across different image resolutions, ensuring broad applicability across diverse object detection tasks. The method's adaptability to changing image conditions and contextual factors makes it particularly effective for complex environments such as surveillance, autonomous vehicles, and industrial automation.

Specification

The present invention relates to the field of computer vision and artificial intelligence, specifically to object detection systems. It focuses on enhancing the accuracy of object detection models using the You Only Look Once (YOLO) architecture by incorporating a dynamic thresholding
mechanism.
BACKGROUND OF THE INVENTION
Object detection is a critical component of many modem computer vision applications, such as autonomous vehicles, surveillance systems, and image analysis. YOLO is a popular deep learning­based object detection architecture known for its speed and efficiency in real-time applications.
However, one of the challenges in object detection is balancing precision and recall, particularly when dealing with varying object sizes, occlusions, and lighting conditions.
Traditional YOLO models use fixed thresholds for classifying detected objects, which may not be optimal in all scenarios. Fixed thresholds can lead to false positives (misclassifying background noise as objects) or false negatives (failing to detect objects). There is a need for an adaptive approach that can dynamically adjust detection thresholds based on the context of the input image to improve
detection accuracy.
SUMMARY OF THE INVENTION
The invention provides a high-precision object detection method based on the YOLO architecture, enhanced by a dynamic thresholding mechanism. The method adjusts detection thresholds in real­ time based on image characteristics such as brightness, contrast, and object density. This dynamic adjustment improves the model's ability to accurately detect objects under varying conditions, reducing false positives and negatives.

Documents