On the influence of artificially distorted images in firearm detection performance using deep learning

Area of Science:

• Computer Vision
• Machine Learning
• Security Technology

Background:

• Automatic firearm detection is crucial for public safety but is hindered by real-world image variations.
• Current deep learning models show promise but their performance under practical image distortions requires thorough investigation.

Purpose of the Study:

• To quantitatively assess the impact of specific image distortions on the performance of deep learning-based firearm detection systems.
• To identify which distortions most significantly degrade detection accuracy when models are trained on clean data.

Main Methods:

• A standard YOLOv5 network was employed for object detection.
• The network was trained on images without distortions and tested on images subjected to various degradations: impulse noise, blurring, darkening, shrinking, and occlusions.

Main Results:

• The study experimentally quantified the performance degradation caused by each distortion type.
• Increased levels of impulse (salt-and-pepper) noise were found to be the most detrimental distortion, significantly impacting detection accuracy.

Conclusions:

• Real-world image conditions, particularly impulse noise, pose a significant challenge to the robustness of current firearm detection systems.
• Further research is needed to develop detection models resilient to these common image degradations for reliable real-world application.