Discrete wavelet transform based steam detection with Adaboost

Steam can cause occlusion in many object detection applications, such as the real problem of large lump detection (LLD) in oil sands mining which motivated our work. In this paper, we propose a general method to overcome this steam detection problem. The existing steam detection methods feasible for our application generally extract features from the transformed input image first and then feed them to a classifier in a completely independent step. In these methods, the step of feature extraction is usually cumbersome and application-dependent. Therefore, we propose a new steam detection method by feeding directly the transformed image to an Adaboost classifier. By doing so, we discard the considerable computational load normally dedicated to feature extraction and benefit from the accuracy of the proper classifier built by Adaboost. Finally, experiments on steam and smoke data sets demonstrate that the proposed steam detection method outperforms the competing methods when taking both efficiency and accuracy into account.