اكتشاف وتمييز الكائنات في الصور منخفضة الإضاءة باستخدام تقنيات التعلم الآلي

Object Detection in Low Light Images Using Deep Learning Techniques

ABstract

Object detection in dark environments has unique characteristics, including poor visibility, shadows, and excessive light reflections. All these factors contribute to reduced image quality, making detection difficult. These obstacles are particularly crucial for applications such as monitoring, autonomous systems, and security, where effective object detection is a requirement for any operations. The challenge is further aggravated by the ever-changing combinations of low-light environmental conditions and the complexity of image features.
This Two-Stage Framework investigates the potential of employing new approaches to enhance images, in conjunction with object detection methods designed for low-light conditions. It focuses on both classical techniques of image processing, such as histogram equalization and noise reduction, as well as deep learning-based models used for object detection. The combination of these techniques aims to achieve better and more reliable results in different types of low-light scenarios. Image processing techniques are applied on the Exdark (Exclusively Dark) dataset to enhance image quality and improve feature visibility.
To validate the effectiveness of the proposed enhancement method, two different deep learning object detection models were employed: a one-stage detector, YOLOv9, and a two-stage detector, Faster R-CNN. The use of these two distinct architectures allows for a comprehensive evaluation of the proposed method across different detection paradigms. Experimental results showed that YOLOv9 achieved the highest performance, with an object detection accuracy of 96% mAP@50, while Faster R-CNN also demonstrated strong results with 88% mAP@50. These outcomes confirm the robustness of the proposed method using two different models.
This Two-Stage Framework focuses more on the development of powerful and dynamic computer systems that are capable of functioning in conditions of low light. This work directly tackles the pervasive problem of accurate object detection in practical situations.