Building Detection of High-Resolution Images based on Global-Local Feature Fusion and Adaptive Anchor

Deep convolutional neural networks (CNNs) based methods have achieved great success in object detection. The accessibility to high resolution images increased demands for effective analyses. However, current approaches either downsampling or cropping for separately processing high resolution images can hardly provide sufficient structural or semantic information, which makes small object detection a challenging task. Aiming to these issues, a multi-branch feature fusion network based on the two-stage detector is proposed, in which a global branch and a local branch takes downsampled images and cropped patches as inputs and extracts contextual and structural features respectively. Then the extracted features are deeply and effectively merged into enhanced feature maps in fusion branch. Preliminary experiments verify that the multi-branch network is effective to preserve both contextual and structural information. However, the anchor-based branches rely on a set of pre-defined anchor scales, which make detection performances sensitive to the hyper-parameters related to anchor boxes and lead to feature inconsistency as base anchor scales for global branch and local branch are favored by objects of downsampled and unchanged respectively. To mitigate feature inconsistency, an improved guided anchoring strategy for the multi-branch network is introduced for adaptive anchor generation, which makes the fused feature map adaptive to fit the shape of anchors. A set of experiments are implemented from the perspective of global inference and local inference to demonstrate the effectiveness of the proposed method on the public XVIEW dataset. The extensive results show that the proposed method is significantly better than other state-of-the-art detectors in the task of building detection.