面向天基监视的红外弱小飞行目标识别算法

针对当前红外弱小飞行目标特征不明显、背景干扰大等问题，提出了一种基于深度学习的红外弱小目标识别算法。检测框架以YOLOv4模型为基础，通过使用K-means++算法对训练集的候选框进行聚类处理，在初始大小的选取上放弃随机生成初始点的方式，在样本集里选取某一个样本作为初始中心使锚框（anchor）大小的选取更加合理。在模型结构中引入卷积注意力模块，使算法模型计算资源分配更合理，对红外弱小飞行目标的特征信息更加敏感。改进空间金字塔池化模块，使用平均池化可以更多保留图像的原始信息，降低天基成像中的噪点与坏点的影响。仿真实验表明采用K-means++计算Anchor大小时准确率可以达到80.13%，在加入了SPP和CBAM模块后之后在测试集上算法识别准确率达到了83.3%，经过对模型的修改有效提升了对红外弱小飞行目标识别的准确率。

Infrared dim small flying target recognition algorithm for space-based surveillance

Aiming at the problems such as the infrared weak flying targets being not significant，and the difficulty of designing the artificial feature extractor，an infrared dim flying target recognition algorithm was proposed based on deep learning.Similarly，background interference could also be a major problem.Therefore，a detection framework was established to deal with the mentioned problems based on the YOLOv4 model.K-means++ algorithm was used to cluster the candidate frames of the training set.This framework was demonstrated to select the sufficient size of anchor，and it was more reasonable to select one specific sample from the database as the initial center other than picking the initial point randomly.At the same time，the convolutional attention module was introduced into the framework，which not only made a more reasonabe allocation of algorithm resources but also made the information more sensitive in detecting the infrared weak flying targets.Since the spatial pyramid pooling module was well enhanced，more original information about the photograph could be retained by using the average pooling method，which could reduce the effect caused by noise and dead pixels within the space-based imaging process.The experiential simulation illustrates the accuracy can reach 80.13% when calculating the anchor size based on K-means++ modulus.The recognition of the algorithm could even reach 83.3% if adding SPP and CBAM modules to the test set.Furthermore，the accuracy of detecting the infrared weak small flying targets also gains an exceptional improvement after the adjustment on modulus.