A Fine-Grained Object Detection Framework Based on Fixed ROI Masking and Feature Optimization in Optical Remote Sensing Images
Zhang Xiaohan, Yafei Lv, Bi Aipeng, Zhao Jianming, Yao Libo · 2021 International Conference on Control, Automation and Information Sciences (ICCAIS) · 2021
Fine-grained object detection in images aims at providing both precise positions and fine-grained categories of objects of interest, which is a challenging but meaningful task. In this paper, fine-grained object detection in optical remote sensing images is explored. To address the inefficiency of previous methods, firstly an end-to-end fine-grained detection framework with a localization branch and a classification branch is proposed. Convolutional Neural Network (CNN) is used for feature extraction and the shallow features are shared by the two branches, while high-level features in the two branches are learned independently. Then a novel fixed ROI Masking method is designed to extract the features of ROIs with fixed size and a grouped spatial attention mechanism is proposed to enhance the ROI feature maps in classification branch by means of teaching the network “where to see”. Experiments with Google Earth data prove the effectiveness of the proposed method.