EVALUATING THE ROBUSTNESS OF YOLO OBJECT DETECTION ALGORITHM IN TERMS OF DETECTING OBJECTS IN NOISY ENVIRONMENT

Our daily lives are impacted by object detection in many ways, such as automobile driving, traffic control, medical fields, etc. Over the past few years, deep learning techniques have been widely used for object detection. Several powerful models have been developed over the past decade for this purpose. The YOLO architecture is one of the most important cutting-edge approaches to object detection. Researchers have used YOLO in their object detection tasks and obtained promising results. Since the YOLO algorithm can be used as an object detector in critical domains, it should provide a quite high accuracy both in noisy and noise-free environments. Consequently, in this study, we aim to carry out an experimental study to test the robustness of the YOLO v5 object detection algorithm when applied to noisy environments. To this end, four case studies have been conducted to evaluate this algorithm's ability to detect objects in noisy images. Specifically, four datasets have been created by injecting an original quality image dataset with different ratios of Gaussian noise. The YOLO v5 algorithm has been trained and tested using the original high-quality dataset. Then, the trained YOLO algorithm has been tested using the created noisy image datasets to monitor the changes in its performance in proportion to the injected Gaussian noise ratio. To our knowledge, this type of performance evaluation study did not conduct before in the literature. Furthermore, there are no such noisy image datasets have been shared before for conducting these types of studies. The obtained results showed that the YOLO algorithm failed to handle the noisy images efficiently besides degrading its performance in proportion to noise rates.