Technological advancements have normalized the usage of unmanned aerial vehicles (UAVs) in every sector, spanning from military to commercial but they also pose serious security concerns due to their enhanced functionalities and easy access to private and highly secured areas. Several instances related to UAVs have raised security concerns, leading to UAV detection research studies. Visual techniques are widely adopted for UAV detection, but they perform poorly at night, in complex backgrounds, and in adverse weather conditions. Therefore, a robust night vision-based drone detection system is required to that could efficiently tackle this problem. Infrared cameras are increasingly used for nighttime surveillance due to their wide applications in night vision equipment. This paper uses a deep learning-based TinyFeatureNet (TF-Net), which is an improved version of YOLOv5s, to accurately detect UAVs during the night using infrared (IR) images. In the proposed TF-Net, we introduce architectural changes in the neck and backbone of the YOLOv5s. We also simulated four different YOLOv5 models (s,m,n,l) and proposed TF-Net for a fair comparison. The results showed better performance for the proposed TF-Net in terms of precision, IoU, GFLOPS, model size, and FPS compared to the YOLOv5s. TF-Net yielded the best results with 95.7\% precision, 84\% mAp, and 44.8\% $IoU$.

基于训练后辍学的方法实现了高稀疏性，并且是解释与计算成本和神经网络架构中过度拟合的问题的良好方法。相反，初始化修剪仍然远远落后。当涉及到网络的计算成本时，初始化修剪更有效。此外，它可以处理过度拟合以及培训后辍学。在对上述原因的认可中，本文提出了两种初始化时修剪的方法。目标是在保持性能的同时获得更高的稀疏性。 1）K-starts，从初始化时k随机p-sparse矩阵开始。在前几个时期，网络然后确定了这些P-Sparse矩阵的“优胜者”，以尝试找到“彩票” P-SPARSE网络。进化算法如何找到最好的个体来采用这种方法。根据神经网络体系结构，健身标准可以基于网络权重的大小，梯度积累的幅度或两者的组合。 2）耗散梯度方法，目的是消除在前几个时期内保持其初始值的一部分的权重。尽管它们的幅度最佳地保留了网络的性能，但以这种方式去除权重。相反，该方法还需要最多的时期才能达到更高的稀疏性。 3）耗散梯度和KSTART的组合始终优于方法和随机辍学。使用提供的相关方法的好处是：1）他们不需要对分类任务的特定知识，固定辍学阈值或正则化参数2）模型的重新训练既不是必要的，也不影响P-SPARSE网络的性能。

大型预先训练的语言模型已经显示了几次拍摄学习的承诺，只提供了几个任务特定示例给出了基于文本的任务。款式将很快解决到目前为止为人类研究助理保留的分类任务吗？现有的基准标记不设计用于衡量应用设置的进度，因此不要直接回答这个问题。 RAFT基准（现实世界注释的少量拍摄任务）侧重于自然发生的任务，并使用镜像部署的评估设置。 RAFT的基线评估揭示了当前技术斗争的地区：推理在许多班级的长篇文章和任务上。人类基线表明，非专家人类难以反映出一些分类任务，反映了现实世界的价值有时依赖于域名专业知识。甚至非专业人类基线F1分数超过GPT-3平均为0.11。 RAFT DataSets和排行榜将跟踪哪些模型改进在https://raft.elict.org中转化为现实世界的优势。