Dong Yan, Liu Yundong, Cheng Yuhua, Gao Guangshuai, Chen Kai, Li Chunlei
School of Automation Engineering, University of Electronic Science and Technology of China, Chengdu, China.
School of Electronic and Information Engineering, Zhongyuan University of Technology, ZhengZhou, China.
PeerJ Comput Sci. 2024 Jul 29;10:e2199. doi: 10.7717/peerj-cs.2199. eCollection 2024.
Accurate localization of objects of interest in remote sensing images (RSIs) is of great significance for object identification, resource management, decision-making and disaster relief response. However, many difficulties, like complex backgrounds, dense target quantities, large-scale variations, and small-scale objects, which make the detection accuracy unsatisfactory. To improve the detection accuracy, we propose an Adaptive Adjacent Context Negotiation Network (ACN-Net). Firstly, the composite fast Fourier convolution (CFFC) module is given to reduce the information loss of small objects, which is inserted into the backbone network to obtain spectral global context information. Then, the Global Context Information Enhancement (GCIE) module is given to capture and aggregate global spatial features, which is beneficial for locating objects of different scales. Furthermore, to alleviate the aliasing effect caused by the fusion of adjacent feature layers, a novel Adaptive Adjacent Context Negotiation network (ACN) is given to adaptive integration of multi-level features, which consists of local and adjacent branches, with the local branch adaptively highlighting feature information and the adjacent branch introducing global information at the adjacent level to enhance feature representation. In the meantime, considering the variability in the focus of feature layers in different dimensions, learnable weights are applied to the local and adjacent branches for adaptive feature fusion. Finally, extensive experiments are performed in several available public datasets, including DIOR and DOTA-v1.0. Experimental studies show that ACN-Net can significantly boost detection performance, with mAP increasing to 74.2% and 79.2%, respectively.
在遥感图像（RSIs）中准确地定位感兴趣的目标对于目标识别、资源管理、决策制定和救灾响应具有重要意义。然而，存在许多困难，如复杂的背景、密集的目标数量、大规模的变化以及小尺度目标，这使得检测精度不尽人意。为了提高检测精度，我们提出了一种自适应相邻上下文协商网络（ACN-Net）。首先，给出复合快速傅里叶卷积（CFFC）模块以减少小目标的信息损失，将其插入主干网络以获得频谱全局上下文信息。然后，给出全局上下文信息增强（GCIE）模块以捕获和聚合全局空间特征，这有利于定位不同尺度的目标。此外，为了减轻相邻特征层融合引起的混叠效应，给出一种新颖的自适应相邻上下文协商网络（ACN）以对多级特征进行自适应集成，其由局部和相邻分支组成，局部分支自适应地突出特征信息，相邻分支在相邻级别引入全局信息以增强特征表示。同时，考虑到不同维度特征层焦点的变异性，将可学习权重应用于局部和相邻分支以进行自适应特征融合。最后，在几个可用的公共数据集上进行了广泛的实验，包括DIOR和DOTA-v1.0。实验研究表明，ACN-Net可以显著提高检测性能，mAP分别提高到74.2%和79.2%。
