A double mapping framework for extraction of shape-invariant features based on multi-scale partitions with AdaBoost for video smoke detection

Abstract

Traditional methods for video smoke detection can easily achieve very low training errors but their generalization performances are not good due to arbitrary shapes of smoke, intra-class variations, occlusions and clutters. To overcome these problems, a double mapping framework is proposed to extract partition based features with AdaBoost. The first mapping is from an original image to block features. A feature vector is presented by concatenating histograms of edge orientation, edge magnitude and Local Binary Pattern (LBP) bit, and densities of edge magnitude, LBP bit, color intensity and saturation. Each component of the feature vector produces a feature image. To obtain shape-invariant features, a detection window is partitioned into a set of small blocks called a partition, and many multi-scale partitions are generated by changing block sizes and partition schemes. The sum of each feature image within each block of each partition is computed to generate block features. The second mapping is from the block features to statistical features. The statistical features of the block features, such as, mean, variance, skewness, kurtosis and Hu moments, are computed on all partitions to form a feature pool. AdaBoost is used to select discriminative shape-invariant features from the feature pool. Experiments show that the proposed method has better generalization performance and less insensitivity to geometry transform than traditional methods.