Attention Region Latent SVM for Image Classification

This paper presents a new method for image classification based on image saliency region. The proposed attention region latent SVM (ARLSVM) is highly distinctive by training in a weakly-supervised manner which without requiring objects position or bounding boxes in training images. We use a latent SVM to model the optimization problem with saliency regions are latent variables. An EM method is proposed to solve the semi-convex optimization problem. Through experiments, our proposed approach performs favourably compared with two well-known algorithms in a benchmark dataset.