Manifold-Regularized Feature Selector for High-Resolution Aerial Photographs Categorization

Recognizing each aerial photo with high-resolution (HR) is a useful technology in image understanding. Herein, a manifold-regularized feature selection (MRFS) is designed to acquire discriminative perceptual features that classify HR aerial images into different categories. Practically, human visual cognition process reflects that, in a scenic picture, the less visually attractive image patches are highly related. Meanwhile, the foreground highly visually attractive image patches are practically unrelated with each other. Following this observation, we in this work propose a multi-layer low-rank paradigm which calculates a succinct set of visually attractive patches in foreground. We sequentially link the above attractive image patches to build a so-called gaze shifting path (GSP). GSP can mimick how humans perceiving different aerial images. Afterward, we formulate a MRFS framework to obtain a subset of high quality features from the entire deep GSP representation. Thereby, an SVM is learned simultaneously. Moreover, the distribution of HR aerial images on the underlying manifold can be maximally preserved during feature selection (FS). To comprehensively evaluate our method, we collect a massive-scale aerial image set containing over 4.87 million high- and low-resolution aerial images. Extensive empirical validations on have shown our algorithm’s efficiency and effectiveness: 1) the testing time cost is 0.8s faster than the second best one to categorize each HR aerial image, 2) the average categorization accuracy is over 4.5% higher than the second best one.