Abstract
Typically, image-based smoke detection is formulated as a frame classification task that aims to automatically assign the captured frames to the predefined "smoke" or "smoke-free" classes. This classification is based on the visual content of the images. In other words, the keystone of such a solution is the choice of the visual descriptor(s) used to encode the visual characteristics of the smoke into numerical vectors. In this paper, we propose to learn a new feature space to represent the visual descriptors extracted from the video frames in an unsupervised manner. This mapping is intended to yield better discrimination between smoke-free images and those showing smoke patterns. The proposed approach is inspired by the linear hyperspectral unmixing techniques. It defines the axes of the new feature space as the vertices of a minimum-volume simplex enclosing all image pixels in the frame. The obtained empirical results prove that the proposed feature mapping approach reinforces the discrimination power of the visual descriptors and produces better smoke detection performance. In addition, the proposed approach exhibits the valuable ability to automatically determine the most relevant visual descriptors.