The Impact of Noise and Brightness on Object Detection Methods

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Deep learning models, particularly pre-trained neural networks, are widely used for object detection and classification in images.
• Image acquisition limitations can introduce sensor noise, such as Gaussian noise, and brightness alterations, negatively impacting model performance.

Purpose of the Study:

• To investigate the impact of common image degradations (noise and brightness) on state-of-the-art object detection models.
• To analyze how object size interacts with these degradations to affect detection performance.

Main Methods:

• Evaluation of several state-of-the-art object detectors, including YOLO and Faster-RCNN.
• Systematic experiments applying various types and levels of Gaussian noise and brightness alterations to test images.

Main Results:

• Noise and brightness variations significantly degrade the performance of object detection models.
• The performance degradation is exacerbated by the size of the objects being detected.
• Object size is a critical factor influencing model robustness to image quality issues.

Conclusions:

• Image quality is a crucial factor for reliable deep learning-based object detection.
• Object size must be considered when assessing the performance and robustness of object detectors in real-world scenarios.
• Further research may focus on developing noise-robust or adaptive object detection methods.