The Construction of Applied Writing Training Network System

This document explains and demonstrates how to construct the applied writing training in network system. This system which is used in network consists of four modules, including task module, structure training module, text training module and study evaluation module. With the advantages of instantaneity, interactivity, authenticity and adequation of learning resources, the applied writing abilities of the students can be improved effectively by using the network system. And the problems during the process of applied writing teaching, for example, the teaching contents lose contract with social life, the textbook contents are lack of innovation, and writing training goes against cognizing system, can be solved effectively.

基于Deformable Transformer和自适应检测头的遥感图像目标检测

针对光学遥感图像目标精准定位困难、分类和定位特征容易存在冲突等问题,提出了一种基于Deformable Transformer和自适应检测头的遥感图像目标检测方法。首先,设计基于特征融合和Deformable Transformer的特征提取网络,其中特征融合模块能丰富卷积神经网络浅层特征的语义信息,Deformable Transformer能对远距离特征建立依赖,可以有效实现对全局语义信息的捕获,提升特征表达能力。其次,构建基于任务学习模块的自适应检测头,在检测头中强化任务感知,能够自动学习与调整分类和定位任务的特征表示,缓解特征冲突。最后,将L1-IoU loss作为定位损失函数,在训练过程中能使模型更准确地衡量候选框与真实框之间的定位误差,从而提高目标定位的准确性。在高分辨率遥感数据集NWPU VHR-10和RSOD上对该方法进行有效性评估,结果显示,与其他方法相比,所提方法具有较为明显的提升效果。