| Popis: |
In the remote sensing field, deep learning-based methods have become mainstream for remote sensing image object detection in recent years. However, traditional methods, such as convolutional neural networks (CNNs), mainly ignore the dependencies between features, failing to capture the spatial relationships and relative positions of objects, which affects the detection performance of dense objects, especially small-size objects. To this end, a high-order feature association network (HOFA-Net) for dense object detection in remote sensing has been proposed to better capture the interdependencies between features of channel and spatial dimensions, yielding more distinguishable features. First, we employ CNNs to learn high-level but low-resolution features of objects. To capture feature interdependencies while retaining crucial information, we design a feature association module based on size adaptation nonlocal. This module partitions the low-resolution and high-level features into local regions and utilizes nonlocal residual connections to capture the local contextual information of objects. In addition, we introduce a high-order feature association (HFA) module designed to learn nonlinear feature correlations and interdependencies within the features. In addition, a covariance normalization acceleration strategy is introduced to accelerate computation. Experimental results on two public remote sensing datasets, including the DOTA dataset and the Tiny Person dataset, demonstrate the superiority and effectiveness of the proposed method through comparative experiments. |
