Image content classification using local context and double thresholding

In this paper we address the problem of image labeling, where the goal is to predict and localize relevant labels from a given set of labels. We have approached this problem by utilizing local feature context with independent label estimation, and a double thresholding method to bring classifier probability outputs to binary solutions. The image is segmented and for each segment a feature vector is computed. The feature vector comprises local features of the respective segment, and the separately averaged features of its left, right, top and bottom neighbors. These feature vectors are tested with SVM classifiers for each class, and a double thresholding on the decision value outputs is applied. Our experiments demonstrate that this approach, although simple, is able to capture context and achieve comparable accuracies with the state-of-the-art methods, without modeling scene-dependent label configurations.