Camouflaged objects are difficult to detect, for both humans and artificial intelligence. But now an AI has been trained to parse objects from their backgrounds.
This could have a variety of聽applications, such as being used聽for search-and-rescue work,聽detecting agricultural pests,聽medical imaging or in military settings.
Detecting camouflaged objects requires visual perception and knowledge. Until now, many AIs have struggled with this task because their algorithms rely on visual cues, such as differences in colour or easily recognisable shapes, to identify objects.
Advertisement
To improve on this, Jianbing Shen at the Inception Institute of Artificial Intelligence in Abu Dhabi in the United Arab Emirates and his colleagues collated a data set of 10,000 photographs to train an AI. The data set includes 5066 images of camouflaged objects, which they have divided into 78 categories, such as 鈥渁mphibian鈥�, 鈥渁quatic鈥� and 鈥渇lying鈥�.
The photographs included both聽naturally camouflaged animals such as fish and insects and examples of artificial camouflage, such as soldiers in聽uniform. Although databases of聽camouflaged objects already exist, this data set is the largest, says Shen.
The team manually labelled each image of a camouflaged object to highlight characteristics such as its shape or whether it was聽partially obstructed by its surrounding environment. They聽then developed an AI called聽SINet and trained it on images from the data set.
The researchers compared SINet to 12 existing algorithms built to detect generic objects. They tested all 13 algorithms using three existing data sets of camouflaged objects. SINet did better than the other 12 at isolating camouflaged objects and identifying their correct shape and nature in both聽the existing and the training聽data sets.
鈥淲ithout any bells and whistles, SINet outperforms various state-of-the-art object detection baselines on all datasets tested, making it a robust, general framework that can help facilitate future research,鈥� the researchers write. They are due to present the work at the CVPR 2020 conference in Seattle, Washington, in June.
The researchers hope the data set and聽algorithm can improve AI鈥檚 ability to recognise camouflaged objects, says Shen.
