Impact of light flickering on object detection accuracy using convolutional neural networks

Machine vision is a key enabling technology in many innovative industries, such as autonomous vehicles, quality inspection, automation and vigilance. Its consistency is crucial for the development of reliable applications. One factor influencing vision systems is lighting, which should be as invariant as possible to allow for consistent image capturing. A threat to consistency is flickering. Light flickering is the high frequency variations in illumination that can happen in some AC-powered lighting systems, such as LED and fluorescent bulbs. The potential impact of flickering in the output confidence of an object detection algorithm is analysed. A neural network algorithm to detect an object under a steady DC-powered light source at 280 lumens was trained. Then, a controlled strobe light was used to assess the confidence of the same model under a pulsating light source at 50 Hz and 100 Hz, mimicking the effects of flickering. The exposure time of the camera was varied across the different frequencies to create 24,000 observations. Statistical analysis proved that flickering can significantly affect the performance of the algorithm, even when not apparent to the human eye. The rationale behind these results is explained, and good practices for setting up similar systems in industrial settings proposed.