提出一种基于深度学习的弱小目标检测方法,该方法基于语义分割任务,利用全卷积递归网络学习复杂背景下弱小目标的特征,并在网络中使用了残差学习和递归操作,具有加速网络优化、模型参数少、深度递归监督和特征重用等特点。将此方法应用在两个真实的图像序列和红外图像测试集上,与三种最新的弱小目标检测方法进行对比,结果显示,在目标增强和背景抑制方面,此方法取得了最好的可视化效果,并在目标检测率、信噪比增益、信杂比增益和背景抑制因子等评价指标上取得了优秀的测试结果。因此,对于不同场景下的红外图像弱小目标检测问题,此方法具有良好的适用性和鲁棒性。

This paper proposes a method for weak target detection based on deep learning. The proposed method based on semantic segmentation uses a fully convolutional recursive network to learn the characteristics of targets in complex backgrounds. Furthermore, it uses residual learning and recursive operation in the network, which exhibits the characteristics of an accelerating network optimization, fewer model parameters, deep recursive supervision, and feature reuse. In two real sequences and other infrared images, the proposed method has achieved the best visual effect in terms of target enhancement and background suppression compared with the three latest detection methods, and it has achieved excellent performance in the probability of detection, signal-to-noise ratio gain, signal-to-clutter ratio gain, and background suppression factor. Therefore, the proposed detection method has good applicability and robustness for infrared dim small target detection in different scenes.