基于深度时空卷积神经网络的点目标检测

由于点目标可用信息少，点目标检测技术是红外搜索与跟踪系统（IRST）中的挑战性难点。基于人工提取特征的传统目标检测，智能化水平低，对点目标检测的难度大。针对此问题，提出一种新的基于深度时空卷积神经网络的点目标检测方法。该方法采用全卷积架构，输入输出尺度相同，可用于处理任意尺度图像。为了提高实时性，卷积分解技术被引入3D时空卷积处理中，将复杂3D时空卷积分解为低复杂度的2D空域卷积和1D时域卷积。根据点目标特点，多权值损失函数被提出，分别采用样本均衡因子和能量均衡因子降低样本不均衡和误差分布不均衡对点目标检测性能的影响。测试结果表明，该方法能够有效抑制复杂背景杂波，并以较低计算量实现点目标检测。

Point target detection based on deep spatial-temporal convolution neural network

Point target detection in Infrared Search and Track (IRST) is a challenging task, due to less information. Traditional methods based on hand-crafted features are hard to finish detection intelligently. A novel deep spatial-temporal convolution neural network is proposed to suppress background and detect point targets. The proposed method is realized based on fully convolution network. So input of arbitrary size can be put into the network and correspondingly-sized output can be obtained. In order to meet the requirement of real time for practical application, the factorized technique is adopted. 3D convolution is decomposed into 2D convolution and 1D convolution, and it leads to signi?cantly less computation. Multi-weighted loss function is designed according to the relation between prediction error and detection performance for point target. Number-balance weight and intensity-balance weight are introduced to deal with the imbalanced sample distribution and imbalanced error distribution. The experimental results show that the proposed method can effectively suppress background clutters, and detect point targets with less runtime.