Mining Mid-Level Visual Elements for Object Detection in High-Resolution Remote Sensing Images

The goal of mining middle-level visual elements is to discover a set of image patches that are representative of and discriminative for a target category. The commonly used mid-level feature representations such as bag-of-visual-words (BOW) models or part-based models in high-resolution remote sensing (HRS) images, seldom consider the discriminability of visual words or parts in object detection. To address this problem, we propose a novel and effective HRS image object detection method based on mid-level visual element representations. First, we employ an iterative procedure that alternates between retraining discriminative classifiers and mining for additional patch instances to discover the discriminative patches, i.e., discriminative mid-level visual elements. Then, a novel mid-level feature representation for an image is constructed based on these visual elements to achieve object detection in HRS images. The experiments on the two HRS image datasets demonstrated the effectiveness of the proposed method compared with several state-of-the-art BOW-based and part-based models.