基于深度卷积神经网络的红外过采样扫描图像点目标检测方法

针对红外过采样扫描成像特点,提出一种基于深度卷积神经网络的红外点目标检测方法.首先,设计回归型深度卷积神经网络以抑制扫描图像杂波背景,该网络不含池化层,输出的背景抑制图像尺寸与输入图像一致;其次,对抑制后的图像进行门限检测,提取候选目标小区域原始数据;最后,将候选目标区域数据依次输入分类型深度卷积神经网络以进一步判别目标、剔除虚警.生成大量过采样训练数据有效训练两个深度网络.结果表明,在不同杂波背景下,该方法在目标信杂比增益、检测概率、虚警概率和运算时间等方面,均优于典型红外小目标检测方法,适用于红外过采样扫描系统的点目标检测.

Point target detection in infrared over-sampling scanning images using deep convolutional neural networks

Aiming at the characteristics of infrared over-sampling scanning imaging, an infrared point target detection method based on DCNN (Deep Convolution Neural Network) is proposed. Firstly, a regressive-type DCNN is designed to suppress the background clutter of the scanning image. The network does not contain any pooling layer, so can input the original image of any size, with the size of output image after clutter suppression in accordance with the input image. Subsequently, the post-suppression image is tested and the original data of candidate target region is extracted. Finally, the candidate target area raw data is input into the classification-type DCNN to further identify the target and remove the false alarm. Meanwhile, a large number of training data of infrared over-sampling scanning images are designed, and two networks are trained effectively. The experimental results show that the proposed method is superior to multiple typical infrared small target detection methods in the target clutter ratio gain, detection probability, false alarm probability and running time under different clutter backgrounds, and is applicable to the point target detection of the infrared oversampling scanning system.