摘要
快速准确地识别遥感图像中小而密集、成像质量不高的目标成为基于深度学习技术的遥感目标检测研究的重点和难点。为解决这类问题,提出结合注意力机制的CenterNet-based目标检测方法,该方法利用ResNet-50网络进行基础特征提取;在主干输出处引入改进的通道注意力模块(ECA-NET),削弱非关注点表达的同时,增强关注点的信息通道;阶段性调整学习策略,加快网络收敛速度。在遥感数据集上进行实验,与原CenterNet算法相比精度提高了13%,检测速度达52.61帧/s。实验结果表明,改进的CenterNet-based目标检测方法在保证检测速度的前提下极大保留小目标的特征,从而降低漏检率,提升识别率,有效平衡了遥感目标检测的精度和速度,在实际环境中具有重要的应用价值。
Fast and accurate identification of small and dense targets in low quality remote sensing images has become a focus of research community. A CenterNet-based target detection method that incorporates the attention mechanism is proposed. It utilizes ResNet-50 for basic feature extraction. An improved channel attention module(ECA-NET) is introduced at the backbone output to weaken the expression of non-concerned points, while enhancing the information channel of concerned points. The proposed method adjusts the learning strategy in different stages to accelerate the convergence speed of the model. Experiments are conducted using the remote sensing dataset. The improved CenterNet algorithm increases the accuracy by 13% as compared with the original method, and the detection speed reaches 52.61 frames per second. The experimental results show that CenterNet-based target detection method maximizes the remote sensing target representation capability of CenterNet under the condition of maintaining computational efficiency. It effectively balances the accuracy and computation speed of remote sensing target detection and is of great significance in practical applications.
作者
黄佳琦
范军芳
李蓓蓓
HUANG Jiaqi;FAN Junfang;LI Beibei(Beijing Key Laboratory of High Dynamic Navigation Technology,Beijing Information Science and Technology University,Beijing 100192,China;School of Automation,Beijing Information Science and Technology University,Beijing 100192,China;No.32381 Unit,Beijing 100072,China)
出处
《弹箭与制导学报》
北大核心
2023年第1期24-31,40,共9页
Journal of Projectiles,Rockets,Missiles and Guidance
基金
国家重点研发计划(2020YFC1511705)
国家自然科学基金(61801032)资助。
